Spaces:
Running
Running
results
Browse filesSigned-off-by: Jonathan Bnayahu <bnayahu@il.ibm.com>
results/bluebench/2025-06-21T08-38-27_evaluation_results.json
ADDED
|
@@ -0,0 +1,1283 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-06-21T12:38:23.605149Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=watsonx/meta-llama/llama-3-2-11b-vision-instruct,max_tokens=256",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "watsonx/meta-llama/llama-3-2-11b-vision-instruct",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 256
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.24.0",
|
| 45 |
+
"unitxt_commit_hash": "2bfd4494ec443ef86013e30d31f4860177124476",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"absl-py": "2.3.0",
|
| 55 |
+
"tiktoken": "0.9.0",
|
| 56 |
+
"charset-normalizer": "3.4.2",
|
| 57 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"litellm": "1.72.6.post1",
|
| 61 |
+
"httpcore": "1.0.9",
|
| 62 |
+
"Jinja2": "3.1.6",
|
| 63 |
+
"jsonschema-specifications": "2025.4.1",
|
| 64 |
+
"pydantic_core": "2.33.2",
|
| 65 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 66 |
+
"yarl": "1.20.1",
|
| 67 |
+
"openai": "1.88.0",
|
| 68 |
+
"portalocker": "3.2.0",
|
| 69 |
+
"pandas": "2.3.0",
|
| 70 |
+
"multiprocess": "0.70.16",
|
| 71 |
+
"jsonschema": "4.24.0",
|
| 72 |
+
"unitxt": "1.24.0",
|
| 73 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 74 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 75 |
+
"pydantic": "2.11.7",
|
| 76 |
+
"async-timeout": "5.0.1",
|
| 77 |
+
"annotated-types": "0.7.0",
|
| 78 |
+
"rouge_score": "0.1.2",
|
| 79 |
+
"contourpy": "1.3.2",
|
| 80 |
+
"aiosignal": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"pillow": "11.2.1",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"pyarrow": "20.0.0",
|
| 87 |
+
"h11": "0.16.0",
|
| 88 |
+
"zipp": "3.19.2",
|
| 89 |
+
"tzdata": "2025.2",
|
| 90 |
+
"bert-score": "0.3.13",
|
| 91 |
+
"setuptools": "80.9.0",
|
| 92 |
+
"referencing": "0.36.2",
|
| 93 |
+
"sacrebleu": "2.5.1",
|
| 94 |
+
"filelock": "3.18.0",
|
| 95 |
+
"urllib3": "2.5.0",
|
| 96 |
+
"scipy": "1.15.3",
|
| 97 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 98 |
+
"kiwisolver": "1.4.8",
|
| 99 |
+
"networkx": "3.4.2",
|
| 100 |
+
"typing-inspection": "0.4.1",
|
| 101 |
+
"lxml": "5.4.0",
|
| 102 |
+
"sniffio": "1.3.1",
|
| 103 |
+
"scikit-learn": "1.7.0",
|
| 104 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 105 |
+
"pip": "25.1.1",
|
| 106 |
+
"fonttools": "4.58.4",
|
| 107 |
+
"transformers": "4.52.4",
|
| 108 |
+
"datasets": "3.6.0",
|
| 109 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 110 |
+
"cycler": "0.12.1",
|
| 111 |
+
"evaluate": "0.4.3",
|
| 112 |
+
"distro": "1.9.0",
|
| 113 |
+
"idna": "3.10",
|
| 114 |
+
"MarkupSafe": "3.0.2",
|
| 115 |
+
"frozenlist": "1.7.0",
|
| 116 |
+
"pyparsing": "3.2.3",
|
| 117 |
+
"jiter": "0.10.0",
|
| 118 |
+
"importlib_metadata": "8.0.0",
|
| 119 |
+
"packaging": "24.2",
|
| 120 |
+
"psutil": "7.0.0",
|
| 121 |
+
"mecab-ko-dic": "1.0.0",
|
| 122 |
+
"joblib": "1.5.1",
|
| 123 |
+
"fsspec": "2025.3.0",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"tokenizers": "0.21.1",
|
| 126 |
+
"wheel": "0.45.1",
|
| 127 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 128 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 129 |
+
"hf-xet": "1.1.4",
|
| 130 |
+
"propcache": "0.3.2",
|
| 131 |
+
"numpy": "2.2.6",
|
| 132 |
+
"mpmath": "1.3.0",
|
| 133 |
+
"multidict": "6.5.0",
|
| 134 |
+
"conllu": "6.0.0",
|
| 135 |
+
"safetensors": "0.5.3",
|
| 136 |
+
"requests": "2.32.4",
|
| 137 |
+
"regex": "2024.11.6",
|
| 138 |
+
"aiohttp": "3.12.13",
|
| 139 |
+
"tabulate": "0.9.0",
|
| 140 |
+
"certifi": "2025.6.15",
|
| 141 |
+
"accelerate": "1.8.0",
|
| 142 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 143 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 144 |
+
"click": "8.2.1",
|
| 145 |
+
"typing_extensions": "4.12.2",
|
| 146 |
+
"attrs": "25.3.0",
|
| 147 |
+
"exceptiongroup": "1.3.0",
|
| 148 |
+
"tenacity": "9.1.2",
|
| 149 |
+
"pytz": "2025.2",
|
| 150 |
+
"aiohappyeyeballs": "2.6.1",
|
| 151 |
+
"python-dateutil": "2.9.0.post0",
|
| 152 |
+
"torch": "2.7.1",
|
| 153 |
+
"python-dotenv": "1.1.0",
|
| 154 |
+
"httpx": "0.28.1",
|
| 155 |
+
"matplotlib": "3.10.3",
|
| 156 |
+
"xxhash": "3.5.0",
|
| 157 |
+
"PyYAML": "6.0.2",
|
| 158 |
+
"huggingface-hub": "0.33.0",
|
| 159 |
+
"colorama": "0.4.6",
|
| 160 |
+
"rpds-py": "0.25.1",
|
| 161 |
+
"threadpoolctl": "3.6.0",
|
| 162 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 0.7555555555555555,
|
| 180 |
+
"accuracy_ci_low": 0.6666666666666666,
|
| 181 |
+
"accuracy_ci_high": 0.8333333333333334,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.7555555555555555,
|
| 184 |
+
"score_ci_high": 0.8333333333333334,
|
| 185 |
+
"score_ci_low": 0.6666666666666666,
|
| 186 |
+
"num_of_instances": 90
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 0.8,
|
| 190 |
+
"accuracy_ci_low": 0.7111111111111111,
|
| 191 |
+
"accuracy_ci_high": 0.8666666666666667,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 0.8,
|
| 194 |
+
"score_ci_high": 0.8666666666666667,
|
| 195 |
+
"score_ci_low": 0.7111111111111111,
|
| 196 |
+
"num_of_instances": 90
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 0.9,
|
| 200 |
+
"accuracy_ci_low": 0.8333333333333334,
|
| 201 |
+
"accuracy_ci_high": 0.9555555555555556,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 0.9,
|
| 204 |
+
"score_ci_high": 0.9555555555555556,
|
| 205 |
+
"score_ci_low": 0.8333333333333334,
|
| 206 |
+
"num_of_instances": 90
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 0.8,
|
| 210 |
+
"accuracy_ci_low": 0.7222222222222222,
|
| 211 |
+
"accuracy_ci_high": 0.8777777777777778,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 0.8,
|
| 214 |
+
"score_ci_high": 0.8777777777777778,
|
| 215 |
+
"score_ci_low": 0.7222222222222222,
|
| 216 |
+
"num_of_instances": 90
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 0.8333333333333334,
|
| 220 |
+
"accuracy_ci_low": 0.7444444444444445,
|
| 221 |
+
"accuracy_ci_high": 0.9,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 0.8333333333333334,
|
| 224 |
+
"score_ci_high": 0.9,
|
| 225 |
+
"score_ci_low": 0.7444444444444445,
|
| 226 |
+
"num_of_instances": 90
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 0.9,
|
| 230 |
+
"accuracy_ci_low": 0.8277783447952625,
|
| 231 |
+
"accuracy_ci_high": 0.9448695638574703,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 0.9,
|
| 234 |
+
"score_ci_high": 0.9448695638574703,
|
| 235 |
+
"score_ci_low": 0.8277783447952625,
|
| 236 |
+
"num_of_instances": 90
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 0.9777777777777777,
|
| 240 |
+
"accuracy_ci_low": 0.9222222222222223,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 0.9777777777777777,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 0.9222222222222223,
|
| 246 |
+
"num_of_instances": 90
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 0.8555555555555555,
|
| 250 |
+
"accuracy_ci_low": 0.7777777777777778,
|
| 251 |
+
"accuracy_ci_high": 0.9222222222222223,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 0.8555555555555555,
|
| 254 |
+
"score_ci_high": 0.9222222222222223,
|
| 255 |
+
"score_ci_low": 0.7777777777777778,
|
| 256 |
+
"num_of_instances": 90
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.9,
|
| 260 |
+
"accuracy_ci_low": 0.822517354780987,
|
| 261 |
+
"accuracy_ci_high": 0.9444444444444444,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.9,
|
| 264 |
+
"score_ci_high": 0.9444444444444444,
|
| 265 |
+
"score_ci_low": 0.822517354780987,
|
| 266 |
+
"num_of_instances": 90
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.8,
|
| 270 |
+
"accuracy_ci_low": 0.7111111111111111,
|
| 271 |
+
"accuracy_ci_high": 0.8777777777777778,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.8,
|
| 274 |
+
"score_ci_high": 0.8777777777777778,
|
| 275 |
+
"score_ci_low": 0.7111111111111111,
|
| 276 |
+
"num_of_instances": 90
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 0.8888888888888888,
|
| 280 |
+
"accuracy_ci_low": 0.8111111111111111,
|
| 281 |
+
"accuracy_ci_high": 0.9444444444444444,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 0.8888888888888888,
|
| 284 |
+
"score_ci_high": 0.9444444444444444,
|
| 285 |
+
"score_ci_low": 0.8111111111111111,
|
| 286 |
+
"num_of_instances": 90
|
| 287 |
+
},
|
| 288 |
+
"score": 0.8555555555555555,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 990
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 500,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.5,
|
| 296 |
+
"score": 0.5,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.5,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 500
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 1000,
|
| 306 |
+
"f1_Person": 0.5,
|
| 307 |
+
"f1_Organization": 0.3540372670807453,
|
| 308 |
+
"f1_Location": 0.35390946502057613,
|
| 309 |
+
"f1_macro": 0.40264891070044045,
|
| 310 |
+
"recall_macro": 0.3613703019405527,
|
| 311 |
+
"precision_macro": 0.46075111602245217,
|
| 312 |
+
"in_classes_support": 0.7734806629834254,
|
| 313 |
+
"f1_micro": 0.3651685393258427,
|
| 314 |
+
"recall_micro": 0.37142857142857144,
|
| 315 |
+
"precision_micro": 0.35911602209944754,
|
| 316 |
+
"score": 0.3651685393258427,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.3095577857641209,
|
| 319 |
+
"score_ci_high": 0.42280766665241554,
|
| 320 |
+
"f1_micro_ci_low": 0.3095577857641209,
|
| 321 |
+
"f1_micro_ci_high": 0.42280766665241554
|
| 322 |
+
},
|
| 323 |
+
"score": 0.3651685393258427,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 1000
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.5492957746478874,
|
| 330 |
+
"accuracy_ci_low": 0.4445856612942033,
|
| 331 |
+
"accuracy_ci_high": 0.6619718309859155,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.5492957746478874,
|
| 334 |
+
"score_ci_high": 0.6619718309859155,
|
| 335 |
+
"score_ci_low": 0.4445856612942033,
|
| 336 |
+
"num_of_instances": 71
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.2676056338028169,
|
| 340 |
+
"accuracy_ci_low": 0.16901408450704225,
|
| 341 |
+
"accuracy_ci_high": 0.38028169014084506,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.2676056338028169,
|
| 344 |
+
"score_ci_high": 0.38028169014084506,
|
| 345 |
+
"score_ci_low": 0.16901408450704225,
|
| 346 |
+
"num_of_instances": 71
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.28169014084507044,
|
| 350 |
+
"accuracy_ci_low": 0.18309859154929578,
|
| 351 |
+
"accuracy_ci_high": 0.39436619718309857,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.28169014084507044,
|
| 354 |
+
"score_ci_high": 0.39436619718309857,
|
| 355 |
+
"score_ci_low": 0.18309859154929578,
|
| 356 |
+
"num_of_instances": 71
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.39436619718309857,
|
| 360 |
+
"accuracy_ci_low": 0.28169014084507044,
|
| 361 |
+
"accuracy_ci_high": 0.5070422535211268,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.39436619718309857,
|
| 364 |
+
"score_ci_high": 0.5070422535211268,
|
| 365 |
+
"score_ci_low": 0.28169014084507044,
|
| 366 |
+
"num_of_instances": 71
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.4788732394366197,
|
| 370 |
+
"accuracy_ci_low": 0.36619718309859156,
|
| 371 |
+
"accuracy_ci_high": 0.5915492957746479,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.4788732394366197,
|
| 374 |
+
"score_ci_high": 0.5915492957746479,
|
| 375 |
+
"score_ci_low": 0.36619718309859156,
|
| 376 |
+
"num_of_instances": 71
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.23943661971830985,
|
| 380 |
+
"accuracy_ci_low": 0.14084507042253522,
|
| 381 |
+
"accuracy_ci_high": 0.3380281690140845,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.23943661971830985,
|
| 384 |
+
"score_ci_high": 0.3380281690140845,
|
| 385 |
+
"score_ci_low": 0.14084507042253522,
|
| 386 |
+
"num_of_instances": 71
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.43661971830985913,
|
| 390 |
+
"accuracy_ci_low": 0.323943661971831,
|
| 391 |
+
"accuracy_ci_high": 0.5492957746478874,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.43661971830985913,
|
| 394 |
+
"score_ci_high": 0.5492957746478874,
|
| 395 |
+
"score_ci_low": 0.323943661971831,
|
| 396 |
+
"num_of_instances": 71
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.5070422535211268,
|
| 400 |
+
"accuracy_ci_low": 0.39436619718309857,
|
| 401 |
+
"accuracy_ci_high": 0.6197183098591549,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.5070422535211268,
|
| 404 |
+
"score_ci_high": 0.6197183098591549,
|
| 405 |
+
"score_ci_low": 0.39436619718309857,
|
| 406 |
+
"num_of_instances": 71
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.28169014084507044,
|
| 410 |
+
"accuracy_ci_low": 0.18309859154929578,
|
| 411 |
+
"accuracy_ci_high": 0.39436619718309857,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.28169014084507044,
|
| 414 |
+
"score_ci_high": 0.39436619718309857,
|
| 415 |
+
"score_ci_low": 0.18309859154929578,
|
| 416 |
+
"num_of_instances": 71
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.14084507042253522,
|
| 420 |
+
"accuracy_ci_low": 0.07042253521126761,
|
| 421 |
+
"accuracy_ci_high": 0.25119134125976145,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.14084507042253522,
|
| 424 |
+
"score_ci_high": 0.25119134125976145,
|
| 425 |
+
"score_ci_low": 0.07042253521126761,
|
| 426 |
+
"num_of_instances": 71
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.2112676056338028,
|
| 430 |
+
"accuracy_ci_low": 0.1267605633802817,
|
| 431 |
+
"accuracy_ci_high": 0.323943661971831,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.2112676056338028,
|
| 434 |
+
"score_ci_high": 0.323943661971831,
|
| 435 |
+
"score_ci_low": 0.1267605633802817,
|
| 436 |
+
"num_of_instances": 71
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.43661971830985913,
|
| 440 |
+
"accuracy_ci_low": 0.323943661971831,
|
| 441 |
+
"accuracy_ci_high": 0.5633802816901409,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.43661971830985913,
|
| 444 |
+
"score_ci_high": 0.5633802816901409,
|
| 445 |
+
"score_ci_low": 0.323943661971831,
|
| 446 |
+
"num_of_instances": 71
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.28169014084507044,
|
| 450 |
+
"accuracy_ci_low": 0.18309859154929578,
|
| 451 |
+
"accuracy_ci_high": 0.39436619718309857,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.28169014084507044,
|
| 454 |
+
"score_ci_high": 0.39436619718309857,
|
| 455 |
+
"score_ci_low": 0.18309859154929578,
|
| 456 |
+
"num_of_instances": 71
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.5633802816901409,
|
| 460 |
+
"accuracy_ci_low": 0.43661971830985913,
|
| 461 |
+
"accuracy_ci_high": 0.676056338028169,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.5633802816901409,
|
| 464 |
+
"score_ci_high": 0.676056338028169,
|
| 465 |
+
"score_ci_low": 0.43661971830985913,
|
| 466 |
+
"num_of_instances": 71
|
| 467 |
+
},
|
| 468 |
+
"score": 0.36217303822937624,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 994
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.5714968851810958,
|
| 475 |
+
"f1_suggestive": 0.38461538461538464,
|
| 476 |
+
"f1_fanciful": 0.4666666666666667,
|
| 477 |
+
"f1_generic": 0.8148148148148148,
|
| 478 |
+
"f1_arbitrary": 0.45454545454545453,
|
| 479 |
+
"f1_descriptive": 0.7368421052631579,
|
| 480 |
+
"f1_macro_ci_low": 0.477080408933172,
|
| 481 |
+
"f1_macro_ci_high": 0.6804913151553064,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.5696969696969697,
|
| 484 |
+
"score_ci_high": 0.6707317073170732,
|
| 485 |
+
"score_ci_low": 0.4662576687116564,
|
| 486 |
+
"num_of_instances": 85,
|
| 487 |
+
"accuracy": 0.5529411764705883,
|
| 488 |
+
"accuracy_ci_low": 0.4470588235294118,
|
| 489 |
+
"accuracy_ci_high": 0.6588235294117647,
|
| 490 |
+
"f1_micro": 0.5696969696969697,
|
| 491 |
+
"f1_micro_ci_low": 0.4662576687116564,
|
| 492 |
+
"f1_micro_ci_high": 0.6707317073170732
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.6068255091609656,
|
| 496 |
+
"f1_no": 0.691358024691358,
|
| 497 |
+
"f1_yes": 0.5222929936305732,
|
| 498 |
+
"f1_macro_ci_low": 0.5410272293620183,
|
| 499 |
+
"f1_macro_ci_high": 0.6785168220322721,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.625,
|
| 502 |
+
"score_ci_high": 0.695,
|
| 503 |
+
"score_ci_low": 0.56,
|
| 504 |
+
"num_of_instances": 200,
|
| 505 |
+
"accuracy": 0.625,
|
| 506 |
+
"accuracy_ci_low": 0.56,
|
| 507 |
+
"accuracy_ci_high": 0.695,
|
| 508 |
+
"f1_micro": 0.625,
|
| 509 |
+
"f1_micro_ci_low": 0.56,
|
| 510 |
+
"f1_micro_ci_high": 0.695
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.341120203240043,
|
| 514 |
+
"f1_conclusion": 0.09523809523809523,
|
| 515 |
+
"f1_analysis": 0.5046728971962616,
|
| 516 |
+
"f1_decree": 0.29411764705882354,
|
| 517 |
+
"f1_issue": 0.21818181818181817,
|
| 518 |
+
"f1_procedural history": 0.34782608695652173,
|
| 519 |
+
"f1_facts": 0.4878048780487805,
|
| 520 |
+
"f1_rule": 0.44,
|
| 521 |
+
"f1_macro_ci_low": 0.28339415146501146,
|
| 522 |
+
"f1_macro_ci_high": 0.41104677447147997,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.36683417085427134,
|
| 525 |
+
"score_ci_high": 0.43609022556390975,
|
| 526 |
+
"score_ci_low": 0.3057644110275689,
|
| 527 |
+
"num_of_instances": 200,
|
| 528 |
+
"accuracy": 0.365,
|
| 529 |
+
"accuracy_ci_low": 0.305,
|
| 530 |
+
"accuracy_ci_high": 0.435,
|
| 531 |
+
"f1_micro": 0.36683417085427134,
|
| 532 |
+
"f1_micro_ci_low": 0.3057644110275689,
|
| 533 |
+
"f1_micro_ci_high": 0.43609022556390975
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.5542857142857143,
|
| 537 |
+
"f1_yes": 0.42857142857142855,
|
| 538 |
+
"f1_no": 0.68,
|
| 539 |
+
"f1_macro_ci_low": 0.48355051877957367,
|
| 540 |
+
"f1_macro_ci_high": 0.6294200433350639,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.5897435897435898,
|
| 543 |
+
"score_ci_high": 0.662674821169658,
|
| 544 |
+
"score_ci_low": 0.5188702892801219,
|
| 545 |
+
"num_of_instances": 200,
|
| 546 |
+
"accuracy": 0.575,
|
| 547 |
+
"accuracy_ci_low": 0.505,
|
| 548 |
+
"accuracy_ci_high": 0.65,
|
| 549 |
+
"f1_micro": 0.5897435897435898,
|
| 550 |
+
"f1_micro_ci_low": 0.5188702892801219,
|
| 551 |
+
"f1_micro_ci_high": 0.662674821169658
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.9155052264808363,
|
| 555 |
+
"f1_yes": 0.9024390243902439,
|
| 556 |
+
"f1_no": 0.9285714285714286,
|
| 557 |
+
"f1_macro_ci_low": 0.8363111281337707,
|
| 558 |
+
"f1_macro_ci_high": 0.9587803500846979,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.9156626506024096,
|
| 561 |
+
"score_ci_high": 0.9585798816568047,
|
| 562 |
+
"score_ci_low": 0.8372630005674523,
|
| 563 |
+
"num_of_instances": 85,
|
| 564 |
+
"accuracy": 0.8941176470588236,
|
| 565 |
+
"accuracy_ci_low": 0.8,
|
| 566 |
+
"accuracy_ci_high": 0.9466354743635627,
|
| 567 |
+
"f1_micro": 0.9156626506024096,
|
| 568 |
+
"f1_micro_ci_low": 0.8372630005674523,
|
| 569 |
+
"f1_micro_ci_high": 0.9585798816568047
|
| 570 |
+
},
|
| 571 |
+
"score": 0.6133874761794481,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 770
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.5088753805047725,
|
| 578 |
+
"f1_cars": 0.7380952380952381,
|
| 579 |
+
"f1_windows x": 0.08333333333333333,
|
| 580 |
+
"f1_computer graphics": 0.5434782608695652,
|
| 581 |
+
"f1_atheism": 0.30434782608695654,
|
| 582 |
+
"f1_religion": 0.1388888888888889,
|
| 583 |
+
"f1_medicine": 0.7733333333333333,
|
| 584 |
+
"f1_christianity": 0.6597938144329897,
|
| 585 |
+
"f1_microsoft windows": 0.39436619718309857,
|
| 586 |
+
"f1_middle east": 0.22641509433962265,
|
| 587 |
+
"f1_politics": 0.3418803418803419,
|
| 588 |
+
"f1_motorcycles": 0.673469387755102,
|
| 589 |
+
"f1_pc hardware": 0.4968944099378882,
|
| 590 |
+
"f1_mac hardware": 0.5617977528089888,
|
| 591 |
+
"f1_electronics": 0.4580152671755725,
|
| 592 |
+
"f1_for sale": 0.4406779661016949,
|
| 593 |
+
"f1_guns": 0.28125,
|
| 594 |
+
"f1_space": 0.7111111111111111,
|
| 595 |
+
"f1_cryptography": 0.6197183098591549,
|
| 596 |
+
"f1_baseball": 0.8617886178861789,
|
| 597 |
+
"f1_hockey": 0.8688524590163934,
|
| 598 |
+
"f1_macro_ci_low": 0.4834761861756795,
|
| 599 |
+
"f1_macro_ci_high": 0.5399453530006973,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.5394515948517068,
|
| 602 |
+
"score_ci_high": 0.571442163952847,
|
| 603 |
+
"score_ci_low": 0.5099380125013132,
|
| 604 |
+
"num_of_instances": 1000,
|
| 605 |
+
"accuracy": 0.482,
|
| 606 |
+
"accuracy_ci_low": 0.45265982508763236,
|
| 607 |
+
"accuracy_ci_high": 0.514,
|
| 608 |
+
"f1_micro": 0.5394515948517068,
|
| 609 |
+
"f1_micro_ci_low": 0.5099380125013132,
|
| 610 |
+
"f1_micro_ci_high": 0.571442163952847
|
| 611 |
+
},
|
| 612 |
+
"score": 0.5394515948517068,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 1000
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.6253924964112395,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.9158878504672897,
|
| 620 |
+
"f1_credit card or prepaid card": 0.625,
|
| 621 |
+
"f1_debt collection": 0.5882352941176471,
|
| 622 |
+
"f1_checking or savings account": 0.75,
|
| 623 |
+
"f1_money transfer or virtual currency or money service": 0.5517241379310345,
|
| 624 |
+
"f1_mortgage": 0.7407407407407407,
|
| 625 |
+
"f1_vehicle loan or lease": 0.5333333333333333,
|
| 626 |
+
"f1_payday loan or title loan or personal loan": 0.1111111111111111,
|
| 627 |
+
"f1_student loan": 0.8125,
|
| 628 |
+
"f1_macro_ci_low": 0.5783727985774465,
|
| 629 |
+
"f1_macro_ci_high": 0.684670784163184,
|
| 630 |
+
"score_name": "f1_micro",
|
| 631 |
+
"score": 0.8322284548699643,
|
| 632 |
+
"score_ci_high": 0.8548513405559257,
|
| 633 |
+
"score_ci_low": 0.809253238376061,
|
| 634 |
+
"num_of_instances": 1000,
|
| 635 |
+
"accuracy": 0.816,
|
| 636 |
+
"accuracy_ci_low": 0.794,
|
| 637 |
+
"accuracy_ci_high": 0.84,
|
| 638 |
+
"f1_micro": 0.8322284548699643,
|
| 639 |
+
"f1_micro_ci_low": 0.809253238376061,
|
| 640 |
+
"f1_micro_ci_high": 0.8548513405559257
|
| 641 |
+
},
|
| 642 |
+
"cfpb_product_watsonx": {
|
| 643 |
+
"f1_macro": 0.6741537589904952,
|
| 644 |
+
"f1_mortgages and loans": 0.7560975609756098,
|
| 645 |
+
"f1_credit card": 0.7589743589743589,
|
| 646 |
+
"f1_debt collection": 0.6829268292682927,
|
| 647 |
+
"f1_credit reporting": 0.7687296416938111,
|
| 648 |
+
"f1_retail banking": 0.40404040404040403,
|
| 649 |
+
"f1_macro_ci_low": 0.6305914704936404,
|
| 650 |
+
"f1_macro_ci_high": 0.7190040244929556,
|
| 651 |
+
"score_name": "f1_micro",
|
| 652 |
+
"score": 0.709278350515464,
|
| 653 |
+
"score_ci_high": 0.7473953353743584,
|
| 654 |
+
"score_ci_low": 0.6666666666666666,
|
| 655 |
+
"num_of_instances": 500,
|
| 656 |
+
"accuracy": 0.688,
|
| 657 |
+
"accuracy_ci_low": 0.644,
|
| 658 |
+
"accuracy_ci_high": 0.726,
|
| 659 |
+
"f1_micro": 0.709278350515464,
|
| 660 |
+
"f1_micro_ci_low": 0.6666666666666666,
|
| 661 |
+
"f1_micro_ci_high": 0.7473953353743584
|
| 662 |
+
},
|
| 663 |
+
"score": 0.7707534026927141,
|
| 664 |
+
"score_name": "subsets_mean",
|
| 665 |
+
"num_of_instances": 1500
|
| 666 |
+
},
|
| 667 |
+
"qa_finance": {
|
| 668 |
+
"fin_qa": {
|
| 669 |
+
"num_of_instances": 1000,
|
| 670 |
+
"execution_accuracy": 0.045,
|
| 671 |
+
"program_accuracy": 0.055,
|
| 672 |
+
"score": 0.055,
|
| 673 |
+
"score_name": "program_accuracy",
|
| 674 |
+
"execution_accuracy_ci_low": 0.033,
|
| 675 |
+
"execution_accuracy_ci_high": 0.058,
|
| 676 |
+
"program_accuracy_ci_low": 0.042,
|
| 677 |
+
"program_accuracy_ci_high": 0.0702228905446176,
|
| 678 |
+
"score_ci_low": 0.042,
|
| 679 |
+
"score_ci_high": 0.0702228905446176
|
| 680 |
+
},
|
| 681 |
+
"score": 0.055,
|
| 682 |
+
"score_name": "subsets_mean",
|
| 683 |
+
"num_of_instances": 1000
|
| 684 |
+
},
|
| 685 |
+
"rag_general": {
|
| 686 |
+
"rag_response_generation_clapnq": {
|
| 687 |
+
"precision": 0.3126916927096655,
|
| 688 |
+
"recall": 0.5297013169878433,
|
| 689 |
+
"f1": 0.3210773896996459,
|
| 690 |
+
"precision_ci_low": 0.2920307369068035,
|
| 691 |
+
"precision_ci_high": 0.33339870629057006,
|
| 692 |
+
"recall_ci_low": 0.512017589843317,
|
| 693 |
+
"recall_ci_high": 0.5445397559503604,
|
| 694 |
+
"f1_ci_low": 0.3051687944542956,
|
| 695 |
+
"f1_ci_high": 0.33801833203554277,
|
| 696 |
+
"score_name": "f1",
|
| 697 |
+
"score": 0.3210773896996459,
|
| 698 |
+
"score_ci_high": 0.33801833203554277,
|
| 699 |
+
"score_ci_low": 0.3051687944542956,
|
| 700 |
+
"num_of_instances": 600,
|
| 701 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.5835817578931649,
|
| 702 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.6631206824382146,
|
| 703 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.5450189588715633,
|
| 704 |
+
"faithfullness_f1_token_overlap": 0.26865346120635386,
|
| 705 |
+
"faithfullness_recall_token_overlap": 0.20704283031640539,
|
| 706 |
+
"faithfullness_precision_token_overlap": 0.5397055221971085,
|
| 707 |
+
"correctness_f1_token_overlap": 0.3210773896996459,
|
| 708 |
+
"correctness_recall_token_overlap": 0.5297013169878433,
|
| 709 |
+
"correctness_precision_token_overlap": 0.3126916927096655
|
| 710 |
+
},
|
| 711 |
+
"score": 0.3210773896996459,
|
| 712 |
+
"score_name": "subsets_mean",
|
| 713 |
+
"num_of_instances": 600
|
| 714 |
+
},
|
| 715 |
+
"reasoning": {
|
| 716 |
+
"hellaswag": {
|
| 717 |
+
"accuracy": 0.475,
|
| 718 |
+
"accuracy_ci_low": 0.4436611532620239,
|
| 719 |
+
"accuracy_ci_high": 0.504,
|
| 720 |
+
"score_name": "accuracy",
|
| 721 |
+
"score": 0.475,
|
| 722 |
+
"score_ci_high": 0.504,
|
| 723 |
+
"score_ci_low": 0.4436611532620239,
|
| 724 |
+
"num_of_instances": 1000
|
| 725 |
+
},
|
| 726 |
+
"openbook_qa": {
|
| 727 |
+
"accuracy": 0.798,
|
| 728 |
+
"accuracy_ci_low": 0.7633491592715252,
|
| 729 |
+
"accuracy_ci_high": 0.832,
|
| 730 |
+
"score_name": "accuracy",
|
| 731 |
+
"score": 0.798,
|
| 732 |
+
"score_ci_high": 0.832,
|
| 733 |
+
"score_ci_low": 0.7633491592715252,
|
| 734 |
+
"num_of_instances": 500
|
| 735 |
+
},
|
| 736 |
+
"score": 0.6365000000000001,
|
| 737 |
+
"score_name": "subsets_mean",
|
| 738 |
+
"num_of_instances": 1500
|
| 739 |
+
},
|
| 740 |
+
"safety": {
|
| 741 |
+
"attaq_500": {
|
| 742 |
+
"severity_critical": 0.0,
|
| 743 |
+
"severity_high": 1.0,
|
| 744 |
+
"severity_medium": 0.0,
|
| 745 |
+
"severity_low": 99.0,
|
| 746 |
+
"category_harmful_info": 0.8553694002921619,
|
| 747 |
+
"category_deception": 0.835255232046951,
|
| 748 |
+
"category_explicit_content": 0.8441415034761331,
|
| 749 |
+
"category_violence": 0.8099433737673931,
|
| 750 |
+
"category_discrimination": 0.770528504891055,
|
| 751 |
+
"category_substance_abuse": 0.8680479208628336,
|
| 752 |
+
"category_pii": 0.8295589322569192,
|
| 753 |
+
"safety": 0.8299434131611555,
|
| 754 |
+
"safety_ci_low": 0.8055372622125345,
|
| 755 |
+
"safety_ci_high": 0.8483225977870763,
|
| 756 |
+
"score_name": "safety",
|
| 757 |
+
"score": 0.8299434131611555,
|
| 758 |
+
"score_ci_high": 0.8483225977870763,
|
| 759 |
+
"score_ci_low": 0.8055372622125345,
|
| 760 |
+
"num_of_instances": 100
|
| 761 |
+
},
|
| 762 |
+
"score": 0.8299434131611555,
|
| 763 |
+
"score_name": "subsets_mean",
|
| 764 |
+
"num_of_instances": 100
|
| 765 |
+
},
|
| 766 |
+
"summarization": {
|
| 767 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 768 |
+
"num_of_instances": 528,
|
| 769 |
+
"rougeLsum": 0.37058702847435154,
|
| 770 |
+
"rouge2": 0.22434285890237585,
|
| 771 |
+
"rougeL": 0.30351270566286986,
|
| 772 |
+
"score": 0.30351270566286986,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rouge1": 0.4255147200867186,
|
| 775 |
+
"rougeLsum_ci_low": 0.36115259966566504,
|
| 776 |
+
"rougeLsum_ci_high": 0.379873984149421,
|
| 777 |
+
"rouge2_ci_low": 0.21612712173196832,
|
| 778 |
+
"rouge2_ci_high": 0.23293503766644144,
|
| 779 |
+
"rougeL_ci_low": 0.2950395615436246,
|
| 780 |
+
"rougeL_ci_high": 0.3118499963672728,
|
| 781 |
+
"score_ci_low": 0.2950395615436246,
|
| 782 |
+
"score_ci_high": 0.3118499963672728,
|
| 783 |
+
"rouge1_ci_low": 0.4147565222299168,
|
| 784 |
+
"rouge1_ci_high": 0.43557795557970264
|
| 785 |
+
},
|
| 786 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 787 |
+
"num_of_instances": 1000,
|
| 788 |
+
"rougeLsum": 0.10643447818455329,
|
| 789 |
+
"rouge2": 0.018171690879646322,
|
| 790 |
+
"rougeL": 0.09269669467501682,
|
| 791 |
+
"score": 0.09269669467501682,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rouge1": 0.12731717955128613,
|
| 794 |
+
"rougeLsum_ci_low": 0.10150863047194304,
|
| 795 |
+
"rougeLsum_ci_high": 0.11036755862721943,
|
| 796 |
+
"rouge2_ci_low": 0.016122174542425746,
|
| 797 |
+
"rouge2_ci_high": 0.02016619357964368,
|
| 798 |
+
"rougeL_ci_low": 0.08847971390285901,
|
| 799 |
+
"rougeL_ci_high": 0.09604727671370761,
|
| 800 |
+
"score_ci_low": 0.08847971390285901,
|
| 801 |
+
"score_ci_high": 0.09604727671370761,
|
| 802 |
+
"rouge1_ci_low": 0.1208888004165388,
|
| 803 |
+
"rouge1_ci_high": 0.13193160708995647
|
| 804 |
+
},
|
| 805 |
+
"score": 0.19810470016894333,
|
| 806 |
+
"score_name": "subsets_mean",
|
| 807 |
+
"num_of_instances": 1528
|
| 808 |
+
},
|
| 809 |
+
"translation": {
|
| 810 |
+
"mt_flores_101_ara_eng": {
|
| 811 |
+
"num_of_instances": 66,
|
| 812 |
+
"counts": [
|
| 813 |
+
1212,
|
| 814 |
+
730,
|
| 815 |
+
491,
|
| 816 |
+
347
|
| 817 |
+
],
|
| 818 |
+
"totals": [
|
| 819 |
+
1843,
|
| 820 |
+
1777,
|
| 821 |
+
1711,
|
| 822 |
+
1645
|
| 823 |
+
],
|
| 824 |
+
"precisions": [
|
| 825 |
+
0.6576234400434074,
|
| 826 |
+
0.41080472706809235,
|
| 827 |
+
0.2869666861484512,
|
| 828 |
+
0.2109422492401216
|
| 829 |
+
],
|
| 830 |
+
"bp": 1.0,
|
| 831 |
+
"sys_len": 1843,
|
| 832 |
+
"ref_len": 1734,
|
| 833 |
+
"sacrebleu": 0.3576036500711557,
|
| 834 |
+
"score": 0.3576036500711557,
|
| 835 |
+
"score_name": "sacrebleu",
|
| 836 |
+
"score_ci_low": 0.30044978362076574,
|
| 837 |
+
"score_ci_high": 0.4017272348853817,
|
| 838 |
+
"sacrebleu_ci_low": 0.30044978362076574,
|
| 839 |
+
"sacrebleu_ci_high": 0.4017272348853817
|
| 840 |
+
},
|
| 841 |
+
"mt_flores_101_deu_eng": {
|
| 842 |
+
"num_of_instances": 66,
|
| 843 |
+
"counts": [
|
| 844 |
+
1262,
|
| 845 |
+
782,
|
| 846 |
+
520,
|
| 847 |
+
359
|
| 848 |
+
],
|
| 849 |
+
"totals": [
|
| 850 |
+
1809,
|
| 851 |
+
1743,
|
| 852 |
+
1677,
|
| 853 |
+
1611
|
| 854 |
+
],
|
| 855 |
+
"precisions": [
|
| 856 |
+
0.6976229961304589,
|
| 857 |
+
0.44865174985656914,
|
| 858 |
+
0.31007751937984496,
|
| 859 |
+
0.2228429546865301
|
| 860 |
+
],
|
| 861 |
+
"bp": 1.0,
|
| 862 |
+
"sys_len": 1809,
|
| 863 |
+
"ref_len": 1734,
|
| 864 |
+
"sacrebleu": 0.3834862844151344,
|
| 865 |
+
"score": 0.3834862844151344,
|
| 866 |
+
"score_name": "sacrebleu",
|
| 867 |
+
"score_ci_low": 0.3416515114541037,
|
| 868 |
+
"score_ci_high": 0.42052556141750186,
|
| 869 |
+
"sacrebleu_ci_low": 0.3416515114541037,
|
| 870 |
+
"sacrebleu_ci_high": 0.42052556141750186
|
| 871 |
+
},
|
| 872 |
+
"mt_flores_101_eng_ara": {
|
| 873 |
+
"num_of_instances": 66,
|
| 874 |
+
"counts": [
|
| 875 |
+
803,
|
| 876 |
+
382,
|
| 877 |
+
189,
|
| 878 |
+
88
|
| 879 |
+
],
|
| 880 |
+
"totals": [
|
| 881 |
+
1643,
|
| 882 |
+
1577,
|
| 883 |
+
1511,
|
| 884 |
+
1445
|
| 885 |
+
],
|
| 886 |
+
"precisions": [
|
| 887 |
+
0.48874010955569086,
|
| 888 |
+
0.2422320862396956,
|
| 889 |
+
0.12508272667107875,
|
| 890 |
+
0.06089965397923876
|
| 891 |
+
],
|
| 892 |
+
"bp": 1.0,
|
| 893 |
+
"sys_len": 1643,
|
| 894 |
+
"ref_len": 1589,
|
| 895 |
+
"sacrebleu": 0.17329277411747399,
|
| 896 |
+
"score": 0.17329277411747399,
|
| 897 |
+
"score_name": "sacrebleu",
|
| 898 |
+
"score_ci_low": 0.1458100600445785,
|
| 899 |
+
"score_ci_high": 0.20408033058998093,
|
| 900 |
+
"sacrebleu_ci_low": 0.1458100600445785,
|
| 901 |
+
"sacrebleu_ci_high": 0.20408033058998093
|
| 902 |
+
},
|
| 903 |
+
"mt_flores_101_eng_deu": {
|
| 904 |
+
"num_of_instances": 66,
|
| 905 |
+
"counts": [
|
| 906 |
+
1126,
|
| 907 |
+
615,
|
| 908 |
+
367,
|
| 909 |
+
220
|
| 910 |
+
],
|
| 911 |
+
"totals": [
|
| 912 |
+
1842,
|
| 913 |
+
1776,
|
| 914 |
+
1710,
|
| 915 |
+
1644
|
| 916 |
+
],
|
| 917 |
+
"precisions": [
|
| 918 |
+
0.6112920738327905,
|
| 919 |
+
0.34628378378378377,
|
| 920 |
+
0.21461988304093566,
|
| 921 |
+
0.13381995133819952
|
| 922 |
+
],
|
| 923 |
+
"bp": 1.0,
|
| 924 |
+
"sys_len": 1842,
|
| 925 |
+
"ref_len": 1835,
|
| 926 |
+
"sacrebleu": 0.27923376694078983,
|
| 927 |
+
"score": 0.27923376694078983,
|
| 928 |
+
"score_name": "sacrebleu",
|
| 929 |
+
"score_ci_low": 0.24552911017678328,
|
| 930 |
+
"score_ci_high": 0.3170919141844367,
|
| 931 |
+
"sacrebleu_ci_low": 0.24552911017678328,
|
| 932 |
+
"sacrebleu_ci_high": 0.3170919141844367
|
| 933 |
+
},
|
| 934 |
+
"mt_flores_101_eng_fra": {
|
| 935 |
+
"num_of_instances": 66,
|
| 936 |
+
"counts": [
|
| 937 |
+
1462,
|
| 938 |
+
1038,
|
| 939 |
+
798,
|
| 940 |
+
621
|
| 941 |
+
],
|
| 942 |
+
"totals": [
|
| 943 |
+
2034,
|
| 944 |
+
1968,
|
| 945 |
+
1902,
|
| 946 |
+
1836
|
| 947 |
+
],
|
| 948 |
+
"precisions": [
|
| 949 |
+
0.7187807276302851,
|
| 950 |
+
0.5274390243902439,
|
| 951 |
+
0.4195583596214511,
|
| 952 |
+
0.338235294117647
|
| 953 |
+
],
|
| 954 |
+
"bp": 0.9834231034146155,
|
| 955 |
+
"sys_len": 2034,
|
| 956 |
+
"ref_len": 2068,
|
| 957 |
+
"sacrebleu": 0.47362585607622854,
|
| 958 |
+
"score": 0.47362585607622854,
|
| 959 |
+
"score_name": "sacrebleu",
|
| 960 |
+
"score_ci_low": 0.4382806139359095,
|
| 961 |
+
"score_ci_high": 0.5139442284621641,
|
| 962 |
+
"sacrebleu_ci_low": 0.4382806139359095,
|
| 963 |
+
"sacrebleu_ci_high": 0.5139442284621641
|
| 964 |
+
},
|
| 965 |
+
"mt_flores_101_eng_kor": {
|
| 966 |
+
"num_of_instances": 66,
|
| 967 |
+
"counts": [
|
| 968 |
+
1295,
|
| 969 |
+
636,
|
| 970 |
+
344,
|
| 971 |
+
187
|
| 972 |
+
],
|
| 973 |
+
"totals": [
|
| 974 |
+
2472,
|
| 975 |
+
2406,
|
| 976 |
+
2340,
|
| 977 |
+
2274
|
| 978 |
+
],
|
| 979 |
+
"precisions": [
|
| 980 |
+
0.5238673139158576,
|
| 981 |
+
0.2643391521197007,
|
| 982 |
+
0.147008547008547,
|
| 983 |
+
0.0822339489885664
|
| 984 |
+
],
|
| 985 |
+
"bp": 1.0,
|
| 986 |
+
"sys_len": 2472,
|
| 987 |
+
"ref_len": 2235,
|
| 988 |
+
"sacrebleu": 0.20227589502234433,
|
| 989 |
+
"score": 0.20227589502234433,
|
| 990 |
+
"score_name": "sacrebleu",
|
| 991 |
+
"score_ci_low": 0.18147140512741464,
|
| 992 |
+
"score_ci_high": 0.2242208833347133,
|
| 993 |
+
"sacrebleu_ci_low": 0.18147140512741464,
|
| 994 |
+
"sacrebleu_ci_high": 0.2242208833347133
|
| 995 |
+
},
|
| 996 |
+
"mt_flores_101_eng_por": {
|
| 997 |
+
"num_of_instances": 66,
|
| 998 |
+
"counts": [
|
| 999 |
+
1388,
|
| 1000 |
+
960,
|
| 1001 |
+
703,
|
| 1002 |
+
527
|
| 1003 |
+
],
|
| 1004 |
+
"totals": [
|
| 1005 |
+
1933,
|
| 1006 |
+
1867,
|
| 1007 |
+
1801,
|
| 1008 |
+
1735
|
| 1009 |
+
],
|
| 1010 |
+
"precisions": [
|
| 1011 |
+
0.7180548370408691,
|
| 1012 |
+
0.5141938939475094,
|
| 1013 |
+
0.3903387007218212,
|
| 1014 |
+
0.3037463976945245
|
| 1015 |
+
],
|
| 1016 |
+
"bp": 1.0,
|
| 1017 |
+
"sys_len": 1933,
|
| 1018 |
+
"ref_len": 1916,
|
| 1019 |
+
"sacrebleu": 0.4574138588750708,
|
| 1020 |
+
"score": 0.4574138588750708,
|
| 1021 |
+
"score_name": "sacrebleu",
|
| 1022 |
+
"score_ci_low": 0.4084699042815335,
|
| 1023 |
+
"score_ci_high": 0.497105169359069,
|
| 1024 |
+
"sacrebleu_ci_low": 0.4084699042815335,
|
| 1025 |
+
"sacrebleu_ci_high": 0.497105169359069
|
| 1026 |
+
},
|
| 1027 |
+
"mt_flores_101_eng_ron": {
|
| 1028 |
+
"num_of_instances": 66,
|
| 1029 |
+
"counts": [
|
| 1030 |
+
1308,
|
| 1031 |
+
869,
|
| 1032 |
+
605,
|
| 1033 |
+
418
|
| 1034 |
+
],
|
| 1035 |
+
"totals": [
|
| 1036 |
+
1942,
|
| 1037 |
+
1876,
|
| 1038 |
+
1810,
|
| 1039 |
+
1744
|
| 1040 |
+
],
|
| 1041 |
+
"precisions": [
|
| 1042 |
+
0.6735324407826982,
|
| 1043 |
+
0.4632196162046908,
|
| 1044 |
+
0.3342541436464089,
|
| 1045 |
+
0.23967889908256879
|
| 1046 |
+
],
|
| 1047 |
+
"bp": 0.9964019571140578,
|
| 1048 |
+
"sys_len": 1942,
|
| 1049 |
+
"ref_len": 1949,
|
| 1050 |
+
"sacrebleu": 0.3961845974220761,
|
| 1051 |
+
"score": 0.3961845974220761,
|
| 1052 |
+
"score_name": "sacrebleu",
|
| 1053 |
+
"score_ci_low": 0.35334192757696886,
|
| 1054 |
+
"score_ci_high": 0.44028759492619396,
|
| 1055 |
+
"sacrebleu_ci_low": 0.35334192757696886,
|
| 1056 |
+
"sacrebleu_ci_high": 0.44028759492619396
|
| 1057 |
+
},
|
| 1058 |
+
"mt_flores_101_eng_spa": {
|
| 1059 |
+
"num_of_instances": 66,
|
| 1060 |
+
"counts": [
|
| 1061 |
+
1251,
|
| 1062 |
+
677,
|
| 1063 |
+
390,
|
| 1064 |
+
223
|
| 1065 |
+
],
|
| 1066 |
+
"totals": [
|
| 1067 |
+
2028,
|
| 1068 |
+
1962,
|
| 1069 |
+
1896,
|
| 1070 |
+
1830
|
| 1071 |
+
],
|
| 1072 |
+
"precisions": [
|
| 1073 |
+
0.6168639053254438,
|
| 1074 |
+
0.34505606523955146,
|
| 1075 |
+
0.20569620253164558,
|
| 1076 |
+
0.12185792349726776
|
| 1077 |
+
],
|
| 1078 |
+
"bp": 0.96607214307495,
|
| 1079 |
+
"sys_len": 2028,
|
| 1080 |
+
"ref_len": 2098,
|
| 1081 |
+
"sacrebleu": 0.2610954642389454,
|
| 1082 |
+
"score": 0.2610954642389454,
|
| 1083 |
+
"score_name": "sacrebleu",
|
| 1084 |
+
"score_ci_low": 0.22712931414581697,
|
| 1085 |
+
"score_ci_high": 0.288932465685729,
|
| 1086 |
+
"sacrebleu_ci_low": 0.22712931414581697,
|
| 1087 |
+
"sacrebleu_ci_high": 0.288932465685729
|
| 1088 |
+
},
|
| 1089 |
+
"mt_flores_101_fra_eng": {
|
| 1090 |
+
"num_of_instances": 66,
|
| 1091 |
+
"counts": [
|
| 1092 |
+
1304,
|
| 1093 |
+
839,
|
| 1094 |
+
567,
|
| 1095 |
+
393
|
| 1096 |
+
],
|
| 1097 |
+
"totals": [
|
| 1098 |
+
1854,
|
| 1099 |
+
1788,
|
| 1100 |
+
1722,
|
| 1101 |
+
1656
|
| 1102 |
+
],
|
| 1103 |
+
"precisions": [
|
| 1104 |
+
0.703344120819849,
|
| 1105 |
+
0.4692393736017897,
|
| 1106 |
+
0.32926829268292684,
|
| 1107 |
+
0.23731884057971014
|
| 1108 |
+
],
|
| 1109 |
+
"bp": 1.0,
|
| 1110 |
+
"sys_len": 1854,
|
| 1111 |
+
"ref_len": 1734,
|
| 1112 |
+
"sacrebleu": 0.4007385379854121,
|
| 1113 |
+
"score": 0.4007385379854121,
|
| 1114 |
+
"score_name": "sacrebleu",
|
| 1115 |
+
"score_ci_low": 0.36457462981738414,
|
| 1116 |
+
"score_ci_high": 0.4478003676387348,
|
| 1117 |
+
"sacrebleu_ci_low": 0.36457462981738414,
|
| 1118 |
+
"sacrebleu_ci_high": 0.4478003676387348
|
| 1119 |
+
},
|
| 1120 |
+
"mt_flores_101_jpn_eng": {
|
| 1121 |
+
"num_of_instances": 66,
|
| 1122 |
+
"counts": [
|
| 1123 |
+
1070,
|
| 1124 |
+
538,
|
| 1125 |
+
304,
|
| 1126 |
+
175
|
| 1127 |
+
],
|
| 1128 |
+
"totals": [
|
| 1129 |
+
1815,
|
| 1130 |
+
1749,
|
| 1131 |
+
1683,
|
| 1132 |
+
1617
|
| 1133 |
+
],
|
| 1134 |
+
"precisions": [
|
| 1135 |
+
0.5895316804407714,
|
| 1136 |
+
0.3076043453401944,
|
| 1137 |
+
0.1806298276886512,
|
| 1138 |
+
0.10822510822510824
|
| 1139 |
+
],
|
| 1140 |
+
"bp": 1.0,
|
| 1141 |
+
"sys_len": 1815,
|
| 1142 |
+
"ref_len": 1734,
|
| 1143 |
+
"sacrebleu": 0.24400811937037023,
|
| 1144 |
+
"score": 0.24400811937037023,
|
| 1145 |
+
"score_name": "sacrebleu",
|
| 1146 |
+
"score_ci_low": 0.2127585740477001,
|
| 1147 |
+
"score_ci_high": 0.2823993625998216,
|
| 1148 |
+
"sacrebleu_ci_low": 0.2127585740477001,
|
| 1149 |
+
"sacrebleu_ci_high": 0.2823993625998216
|
| 1150 |
+
},
|
| 1151 |
+
"mt_flores_101_kor_eng": {
|
| 1152 |
+
"num_of_instances": 66,
|
| 1153 |
+
"counts": [
|
| 1154 |
+
1045,
|
| 1155 |
+
498,
|
| 1156 |
+
280,
|
| 1157 |
+
169
|
| 1158 |
+
],
|
| 1159 |
+
"totals": [
|
| 1160 |
+
1787,
|
| 1161 |
+
1721,
|
| 1162 |
+
1655,
|
| 1163 |
+
1589
|
| 1164 |
+
],
|
| 1165 |
+
"precisions": [
|
| 1166 |
+
0.5847789591494125,
|
| 1167 |
+
0.2893666472980825,
|
| 1168 |
+
0.1691842900302115,
|
| 1169 |
+
0.10635619886721209
|
| 1170 |
+
],
|
| 1171 |
+
"bp": 1.0,
|
| 1172 |
+
"sys_len": 1787,
|
| 1173 |
+
"ref_len": 1734,
|
| 1174 |
+
"sacrebleu": 0.23490418965892076,
|
| 1175 |
+
"score": 0.23490418965892076,
|
| 1176 |
+
"score_name": "sacrebleu",
|
| 1177 |
+
"score_ci_low": 0.1985487418164707,
|
| 1178 |
+
"score_ci_high": 0.2846511928700541,
|
| 1179 |
+
"sacrebleu_ci_low": 0.1985487418164707,
|
| 1180 |
+
"sacrebleu_ci_high": 0.2846511928700541
|
| 1181 |
+
},
|
| 1182 |
+
"mt_flores_101_por_eng": {
|
| 1183 |
+
"num_of_instances": 66,
|
| 1184 |
+
"counts": [
|
| 1185 |
+
1313,
|
| 1186 |
+
888,
|
| 1187 |
+
630,
|
| 1188 |
+
458
|
| 1189 |
+
],
|
| 1190 |
+
"totals": [
|
| 1191 |
+
1824,
|
| 1192 |
+
1758,
|
| 1193 |
+
1692,
|
| 1194 |
+
1626
|
| 1195 |
+
],
|
| 1196 |
+
"precisions": [
|
| 1197 |
+
0.7198464912280701,
|
| 1198 |
+
0.5051194539249146,
|
| 1199 |
+
0.3723404255319149,
|
| 1200 |
+
0.2816728167281673
|
| 1201 |
+
],
|
| 1202 |
+
"bp": 1.0,
|
| 1203 |
+
"sys_len": 1824,
|
| 1204 |
+
"ref_len": 1734,
|
| 1205 |
+
"sacrebleu": 0.4419058783059329,
|
| 1206 |
+
"score": 0.4419058783059329,
|
| 1207 |
+
"score_name": "sacrebleu",
|
| 1208 |
+
"score_ci_low": 0.39279370658531876,
|
| 1209 |
+
"score_ci_high": 0.49550322461749036,
|
| 1210 |
+
"sacrebleu_ci_low": 0.39279370658531876,
|
| 1211 |
+
"sacrebleu_ci_high": 0.49550322461749036
|
| 1212 |
+
},
|
| 1213 |
+
"mt_flores_101_ron_eng": {
|
| 1214 |
+
"num_of_instances": 66,
|
| 1215 |
+
"counts": [
|
| 1216 |
+
1260,
|
| 1217 |
+
841,
|
| 1218 |
+
596,
|
| 1219 |
+
426
|
| 1220 |
+
],
|
| 1221 |
+
"totals": [
|
| 1222 |
+
1797,
|
| 1223 |
+
1731,
|
| 1224 |
+
1665,
|
| 1225 |
+
1599
|
| 1226 |
+
],
|
| 1227 |
+
"precisions": [
|
| 1228 |
+
0.7011686143572621,
|
| 1229 |
+
0.4858463316002311,
|
| 1230 |
+
0.357957957957958,
|
| 1231 |
+
0.26641651031894936
|
| 1232 |
+
],
|
| 1233 |
+
"bp": 1.0,
|
| 1234 |
+
"sys_len": 1797,
|
| 1235 |
+
"ref_len": 1734,
|
| 1236 |
+
"sacrebleu": 0.4245497977223498,
|
| 1237 |
+
"score": 0.4245497977223498,
|
| 1238 |
+
"score_name": "sacrebleu",
|
| 1239 |
+
"score_ci_low": 0.377958140884616,
|
| 1240 |
+
"score_ci_high": 0.4821428710051954,
|
| 1241 |
+
"sacrebleu_ci_low": 0.377958140884616,
|
| 1242 |
+
"sacrebleu_ci_high": 0.4821428710051954
|
| 1243 |
+
},
|
| 1244 |
+
"mt_flores_101_spa_eng": {
|
| 1245 |
+
"num_of_instances": 66,
|
| 1246 |
+
"counts": [
|
| 1247 |
+
1163,
|
| 1248 |
+
621,
|
| 1249 |
+
368,
|
| 1250 |
+
225
|
| 1251 |
+
],
|
| 1252 |
+
"totals": [
|
| 1253 |
+
1900,
|
| 1254 |
+
1834,
|
| 1255 |
+
1768,
|
| 1256 |
+
1702
|
| 1257 |
+
],
|
| 1258 |
+
"precisions": [
|
| 1259 |
+
0.6121052631578947,
|
| 1260 |
+
0.3386041439476554,
|
| 1261 |
+
0.2081447963800905,
|
| 1262 |
+
0.13219741480611044
|
| 1263 |
+
],
|
| 1264 |
+
"bp": 1.0,
|
| 1265 |
+
"sys_len": 1900,
|
| 1266 |
+
"ref_len": 1734,
|
| 1267 |
+
"sacrebleu": 0.27480632590419196,
|
| 1268 |
+
"score": 0.27480632590419196,
|
| 1269 |
+
"score_name": "sacrebleu",
|
| 1270 |
+
"score_ci_low": 0.24866198798867725,
|
| 1271 |
+
"score_ci_high": 0.33121893225397175,
|
| 1272 |
+
"sacrebleu_ci_low": 0.24866198798867725,
|
| 1273 |
+
"sacrebleu_ci_high": 0.33121893225397175
|
| 1274 |
+
},
|
| 1275 |
+
"score": 0.3336749997417598,
|
| 1276 |
+
"score_name": "subsets_mean",
|
| 1277 |
+
"num_of_instances": 990
|
| 1278 |
+
},
|
| 1279 |
+
"score": 0.49083000843124214,
|
| 1280 |
+
"score_name": "subsets_mean",
|
| 1281 |
+
"num_of_instances": 12472
|
| 1282 |
+
}
|
| 1283 |
+
}
|
results/bluebench/2025-06-21T09-36-54_evaluation_results.json
ADDED
|
@@ -0,0 +1,1283 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-06-21T13:36:49.808623Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=watsonx/meta-llama/llama-3-3-70b-instruct,max_tokens=256",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "watsonx/meta-llama/llama-3-3-70b-instruct",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 256
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.24.0",
|
| 45 |
+
"unitxt_commit_hash": "2bfd4494ec443ef86013e30d31f4860177124476",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"absl-py": "2.3.0",
|
| 55 |
+
"tiktoken": "0.9.0",
|
| 56 |
+
"charset-normalizer": "3.4.2",
|
| 57 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"litellm": "1.72.6.post1",
|
| 61 |
+
"httpcore": "1.0.9",
|
| 62 |
+
"Jinja2": "3.1.6",
|
| 63 |
+
"jsonschema-specifications": "2025.4.1",
|
| 64 |
+
"pydantic_core": "2.33.2",
|
| 65 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 66 |
+
"yarl": "1.20.1",
|
| 67 |
+
"openai": "1.88.0",
|
| 68 |
+
"portalocker": "3.2.0",
|
| 69 |
+
"pandas": "2.3.0",
|
| 70 |
+
"multiprocess": "0.70.16",
|
| 71 |
+
"jsonschema": "4.24.0",
|
| 72 |
+
"unitxt": "1.24.0",
|
| 73 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 74 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 75 |
+
"pydantic": "2.11.7",
|
| 76 |
+
"async-timeout": "5.0.1",
|
| 77 |
+
"annotated-types": "0.7.0",
|
| 78 |
+
"rouge_score": "0.1.2",
|
| 79 |
+
"contourpy": "1.3.2",
|
| 80 |
+
"aiosignal": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"pillow": "11.2.1",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"pyarrow": "20.0.0",
|
| 87 |
+
"h11": "0.16.0",
|
| 88 |
+
"zipp": "3.19.2",
|
| 89 |
+
"tzdata": "2025.2",
|
| 90 |
+
"bert-score": "0.3.13",
|
| 91 |
+
"setuptools": "80.9.0",
|
| 92 |
+
"referencing": "0.36.2",
|
| 93 |
+
"sacrebleu": "2.5.1",
|
| 94 |
+
"filelock": "3.18.0",
|
| 95 |
+
"urllib3": "2.5.0",
|
| 96 |
+
"scipy": "1.15.3",
|
| 97 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 98 |
+
"kiwisolver": "1.4.8",
|
| 99 |
+
"networkx": "3.4.2",
|
| 100 |
+
"typing-inspection": "0.4.1",
|
| 101 |
+
"lxml": "5.4.0",
|
| 102 |
+
"sniffio": "1.3.1",
|
| 103 |
+
"scikit-learn": "1.7.0",
|
| 104 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 105 |
+
"pip": "25.1.1",
|
| 106 |
+
"fonttools": "4.58.4",
|
| 107 |
+
"transformers": "4.52.4",
|
| 108 |
+
"datasets": "3.6.0",
|
| 109 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 110 |
+
"cycler": "0.12.1",
|
| 111 |
+
"evaluate": "0.4.3",
|
| 112 |
+
"distro": "1.9.0",
|
| 113 |
+
"idna": "3.10",
|
| 114 |
+
"MarkupSafe": "3.0.2",
|
| 115 |
+
"frozenlist": "1.7.0",
|
| 116 |
+
"pyparsing": "3.2.3",
|
| 117 |
+
"jiter": "0.10.0",
|
| 118 |
+
"importlib_metadata": "8.0.0",
|
| 119 |
+
"packaging": "24.2",
|
| 120 |
+
"psutil": "7.0.0",
|
| 121 |
+
"mecab-ko-dic": "1.0.0",
|
| 122 |
+
"joblib": "1.5.1",
|
| 123 |
+
"fsspec": "2025.3.0",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"tokenizers": "0.21.1",
|
| 126 |
+
"wheel": "0.45.1",
|
| 127 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 128 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 129 |
+
"hf-xet": "1.1.4",
|
| 130 |
+
"propcache": "0.3.2",
|
| 131 |
+
"numpy": "2.2.6",
|
| 132 |
+
"mpmath": "1.3.0",
|
| 133 |
+
"multidict": "6.5.0",
|
| 134 |
+
"conllu": "6.0.0",
|
| 135 |
+
"safetensors": "0.5.3",
|
| 136 |
+
"requests": "2.32.4",
|
| 137 |
+
"regex": "2024.11.6",
|
| 138 |
+
"aiohttp": "3.12.13",
|
| 139 |
+
"tabulate": "0.9.0",
|
| 140 |
+
"certifi": "2025.6.15",
|
| 141 |
+
"accelerate": "1.8.0",
|
| 142 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 143 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 144 |
+
"click": "8.2.1",
|
| 145 |
+
"typing_extensions": "4.12.2",
|
| 146 |
+
"attrs": "25.3.0",
|
| 147 |
+
"exceptiongroup": "1.3.0",
|
| 148 |
+
"tenacity": "9.1.2",
|
| 149 |
+
"pytz": "2025.2",
|
| 150 |
+
"aiohappyeyeballs": "2.6.1",
|
| 151 |
+
"python-dateutil": "2.9.0.post0",
|
| 152 |
+
"torch": "2.7.1",
|
| 153 |
+
"python-dotenv": "1.1.0",
|
| 154 |
+
"httpx": "0.28.1",
|
| 155 |
+
"matplotlib": "3.10.3",
|
| 156 |
+
"xxhash": "3.5.0",
|
| 157 |
+
"PyYAML": "6.0.2",
|
| 158 |
+
"huggingface-hub": "0.33.0",
|
| 159 |
+
"colorama": "0.4.6",
|
| 160 |
+
"rpds-py": "0.25.1",
|
| 161 |
+
"threadpoolctl": "3.6.0",
|
| 162 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 0.7777777777777778,
|
| 180 |
+
"accuracy_ci_low": 0.6777777777777778,
|
| 181 |
+
"accuracy_ci_high": 0.8555555555555555,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.7777777777777778,
|
| 184 |
+
"score_ci_high": 0.8555555555555555,
|
| 185 |
+
"score_ci_low": 0.6777777777777778,
|
| 186 |
+
"num_of_instances": 90
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 1.0,
|
| 190 |
+
"accuracy_ci_low": 1.0,
|
| 191 |
+
"accuracy_ci_high": 1.0,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 1.0,
|
| 194 |
+
"score_ci_high": 1.0,
|
| 195 |
+
"score_ci_low": 1.0,
|
| 196 |
+
"num_of_instances": 90
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 1.0,
|
| 200 |
+
"accuracy_ci_low": 1.0,
|
| 201 |
+
"accuracy_ci_high": 1.0,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 1.0,
|
| 204 |
+
"score_ci_high": 1.0,
|
| 205 |
+
"score_ci_low": 1.0,
|
| 206 |
+
"num_of_instances": 90
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 1.0,
|
| 210 |
+
"accuracy_ci_low": 1.0,
|
| 211 |
+
"accuracy_ci_high": 1.0,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 1.0,
|
| 214 |
+
"score_ci_high": 1.0,
|
| 215 |
+
"score_ci_low": 1.0,
|
| 216 |
+
"num_of_instances": 90
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 1.0,
|
| 220 |
+
"accuracy_ci_low": 1.0,
|
| 221 |
+
"accuracy_ci_high": 1.0,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 1.0,
|
| 224 |
+
"score_ci_high": 1.0,
|
| 225 |
+
"score_ci_low": 1.0,
|
| 226 |
+
"num_of_instances": 90
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 1.0,
|
| 230 |
+
"accuracy_ci_low": 1.0,
|
| 231 |
+
"accuracy_ci_high": 1.0,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 1.0,
|
| 234 |
+
"score_ci_high": 1.0,
|
| 235 |
+
"score_ci_low": 1.0,
|
| 236 |
+
"num_of_instances": 90
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 1.0,
|
| 240 |
+
"accuracy_ci_low": 1.0,
|
| 241 |
+
"accuracy_ci_high": 1.0,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 1.0,
|
| 244 |
+
"score_ci_high": 1.0,
|
| 245 |
+
"score_ci_low": 1.0,
|
| 246 |
+
"num_of_instances": 90
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 0.9888888888888889,
|
| 250 |
+
"accuracy_ci_low": 0.9407763312346947,
|
| 251 |
+
"accuracy_ci_high": 1.0,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 0.9888888888888889,
|
| 254 |
+
"score_ci_high": 1.0,
|
| 255 |
+
"score_ci_low": 0.9407763312346947,
|
| 256 |
+
"num_of_instances": 90
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.9777777777777777,
|
| 260 |
+
"accuracy_ci_low": 0.9333333333333333,
|
| 261 |
+
"accuracy_ci_high": 1.0,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.9777777777777777,
|
| 264 |
+
"score_ci_high": 1.0,
|
| 265 |
+
"score_ci_low": 0.9333333333333333,
|
| 266 |
+
"num_of_instances": 90
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.9888888888888889,
|
| 270 |
+
"accuracy_ci_low": 0.9283857779145438,
|
| 271 |
+
"accuracy_ci_high": 1.0,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.9888888888888889,
|
| 274 |
+
"score_ci_high": 1.0,
|
| 275 |
+
"score_ci_low": 0.9283857779145438,
|
| 276 |
+
"num_of_instances": 90
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 0.8666666666666667,
|
| 280 |
+
"accuracy_ci_low": 0.7888888888888889,
|
| 281 |
+
"accuracy_ci_high": 0.9333333333333333,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 0.8666666666666667,
|
| 284 |
+
"score_ci_high": 0.9333333333333333,
|
| 285 |
+
"score_ci_low": 0.7888888888888889,
|
| 286 |
+
"num_of_instances": 90
|
| 287 |
+
},
|
| 288 |
+
"score": 0.9636363636363636,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 990
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 500,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.5,
|
| 296 |
+
"score": 0.5,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.5,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 500
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 1000,
|
| 306 |
+
"f1_Person": 0.5744680851063829,
|
| 307 |
+
"f1_Organization": 0.3625,
|
| 308 |
+
"f1_Location": 0.3764705882352941,
|
| 309 |
+
"f1_macro": 0.4378128911138923,
|
| 310 |
+
"recall_macro": 0.39500837147693896,
|
| 311 |
+
"precision_macro": 0.49432322991751926,
|
| 312 |
+
"in_classes_support": 0.6292466765140325,
|
| 313 |
+
"f1_micro": 0.3560732113144759,
|
| 314 |
+
"recall_micro": 0.4076190476190476,
|
| 315 |
+
"precision_micro": 0.31610044313146235,
|
| 316 |
+
"score": 0.3560732113144759,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.31585267626167035,
|
| 319 |
+
"score_ci_high": 0.4082105948099384,
|
| 320 |
+
"f1_micro_ci_low": 0.31585267626167035,
|
| 321 |
+
"f1_micro_ci_high": 0.4082105948099384
|
| 322 |
+
},
|
| 323 |
+
"score": 0.3560732113144759,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 1000
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.704225352112676,
|
| 330 |
+
"accuracy_ci_low": 0.5915492957746479,
|
| 331 |
+
"accuracy_ci_high": 0.8028169014084507,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.704225352112676,
|
| 334 |
+
"score_ci_high": 0.8028169014084507,
|
| 335 |
+
"score_ci_low": 0.5915492957746479,
|
| 336 |
+
"num_of_instances": 71
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.19718309859154928,
|
| 340 |
+
"accuracy_ci_low": 0.11267605633802817,
|
| 341 |
+
"accuracy_ci_high": 0.30985915492957744,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.19718309859154928,
|
| 344 |
+
"score_ci_high": 0.30985915492957744,
|
| 345 |
+
"score_ci_low": 0.11267605633802817,
|
| 346 |
+
"num_of_instances": 71
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.1267605633802817,
|
| 350 |
+
"accuracy_ci_low": 0.056338028169014086,
|
| 351 |
+
"accuracy_ci_high": 0.22535211267605634,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.1267605633802817,
|
| 354 |
+
"score_ci_high": 0.22535211267605634,
|
| 355 |
+
"score_ci_low": 0.056338028169014086,
|
| 356 |
+
"num_of_instances": 71
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.6056338028169014,
|
| 360 |
+
"accuracy_ci_low": 0.49295774647887325,
|
| 361 |
+
"accuracy_ci_high": 0.704225352112676,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.6056338028169014,
|
| 364 |
+
"score_ci_high": 0.704225352112676,
|
| 365 |
+
"score_ci_low": 0.49295774647887325,
|
| 366 |
+
"num_of_instances": 71
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.647887323943662,
|
| 370 |
+
"accuracy_ci_low": 0.5211267605633803,
|
| 371 |
+
"accuracy_ci_high": 0.7605633802816901,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.647887323943662,
|
| 374 |
+
"score_ci_high": 0.7605633802816901,
|
| 375 |
+
"score_ci_low": 0.5211267605633803,
|
| 376 |
+
"num_of_instances": 71
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.323943661971831,
|
| 380 |
+
"accuracy_ci_low": 0.2112676056338028,
|
| 381 |
+
"accuracy_ci_high": 0.43661971830985913,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.323943661971831,
|
| 384 |
+
"score_ci_high": 0.43661971830985913,
|
| 385 |
+
"score_ci_low": 0.2112676056338028,
|
| 386 |
+
"num_of_instances": 71
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5492957746478874,
|
| 390 |
+
"accuracy_ci_low": 0.43661971830985913,
|
| 391 |
+
"accuracy_ci_high": 0.6619718309859155,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5492957746478874,
|
| 394 |
+
"score_ci_high": 0.6619718309859155,
|
| 395 |
+
"score_ci_low": 0.43661971830985913,
|
| 396 |
+
"num_of_instances": 71
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.6901408450704225,
|
| 400 |
+
"accuracy_ci_low": 0.5774647887323944,
|
| 401 |
+
"accuracy_ci_high": 0.7887323943661971,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.6901408450704225,
|
| 404 |
+
"score_ci_high": 0.7887323943661971,
|
| 405 |
+
"score_ci_low": 0.5774647887323944,
|
| 406 |
+
"num_of_instances": 71
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.5774647887323944,
|
| 410 |
+
"accuracy_ci_low": 0.4647887323943662,
|
| 411 |
+
"accuracy_ci_high": 0.6901408450704225,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.5774647887323944,
|
| 414 |
+
"score_ci_high": 0.6901408450704225,
|
| 415 |
+
"score_ci_low": 0.4647887323943662,
|
| 416 |
+
"num_of_instances": 71
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.29577464788732394,
|
| 420 |
+
"accuracy_ci_low": 0.2112676056338028,
|
| 421 |
+
"accuracy_ci_high": 0.415768498221757,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.29577464788732394,
|
| 424 |
+
"score_ci_high": 0.415768498221757,
|
| 425 |
+
"score_ci_low": 0.2112676056338028,
|
| 426 |
+
"num_of_instances": 71
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.49295774647887325,
|
| 430 |
+
"accuracy_ci_low": 0.38028169014084506,
|
| 431 |
+
"accuracy_ci_high": 0.6056338028169014,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.49295774647887325,
|
| 434 |
+
"score_ci_high": 0.6056338028169014,
|
| 435 |
+
"score_ci_low": 0.38028169014084506,
|
| 436 |
+
"num_of_instances": 71
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.5915492957746479,
|
| 440 |
+
"accuracy_ci_low": 0.4647887323943662,
|
| 441 |
+
"accuracy_ci_high": 0.6901408450704225,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.5915492957746479,
|
| 444 |
+
"score_ci_high": 0.6901408450704225,
|
| 445 |
+
"score_ci_low": 0.4647887323943662,
|
| 446 |
+
"num_of_instances": 71
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.3380281690140845,
|
| 450 |
+
"accuracy_ci_low": 0.22535211267605634,
|
| 451 |
+
"accuracy_ci_high": 0.4507042253521127,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.3380281690140845,
|
| 454 |
+
"score_ci_high": 0.4507042253521127,
|
| 455 |
+
"score_ci_low": 0.22535211267605634,
|
| 456 |
+
"num_of_instances": 71
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.704225352112676,
|
| 460 |
+
"accuracy_ci_low": 0.5915492957746479,
|
| 461 |
+
"accuracy_ci_high": 0.8028169014084507,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.704225352112676,
|
| 464 |
+
"score_ci_high": 0.8028169014084507,
|
| 465 |
+
"score_ci_low": 0.5915492957746479,
|
| 466 |
+
"num_of_instances": 71
|
| 467 |
+
},
|
| 468 |
+
"score": 0.48893360160965793,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 994
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.6576598179941833,
|
| 475 |
+
"f1_suggestive": 0.5882352941176471,
|
| 476 |
+
"f1_generic": 0.72,
|
| 477 |
+
"f1_descriptive": 0.6818181818181818,
|
| 478 |
+
"f1_fanciful": 0.6666666666666666,
|
| 479 |
+
"f1_arbitrary": 0.631578947368421,
|
| 480 |
+
"f1_macro_ci_low": 0.5539989867710009,
|
| 481 |
+
"f1_macro_ci_high": 0.7630121059181559,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.6547619047619048,
|
| 484 |
+
"score_ci_high": 0.75,
|
| 485 |
+
"score_ci_low": 0.5476190476190477,
|
| 486 |
+
"num_of_instances": 85,
|
| 487 |
+
"accuracy": 0.6470588235294118,
|
| 488 |
+
"accuracy_ci_low": 0.5411764705882353,
|
| 489 |
+
"accuracy_ci_high": 0.7411764705882353,
|
| 490 |
+
"f1_micro": 0.6547619047619048,
|
| 491 |
+
"f1_micro_ci_low": 0.5476190476190477,
|
| 492 |
+
"f1_micro_ci_high": 0.75
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.5644255319148936,
|
| 496 |
+
"f1_no": 0.6808510638297872,
|
| 497 |
+
"f1_yes": 0.448,
|
| 498 |
+
"f1_macro_ci_low": 0.4955233440030641,
|
| 499 |
+
"f1_macro_ci_high": 0.637195719539535,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.6,
|
| 502 |
+
"score_ci_high": 0.6666666666666666,
|
| 503 |
+
"score_ci_low": 0.5292645910446814,
|
| 504 |
+
"num_of_instances": 200,
|
| 505 |
+
"accuracy": 0.54,
|
| 506 |
+
"accuracy_ci_low": 0.47,
|
| 507 |
+
"accuracy_ci_high": 0.605,
|
| 508 |
+
"f1_micro": 0.6,
|
| 509 |
+
"f1_micro_ci_low": 0.5292645910446814,
|
| 510 |
+
"f1_micro_ci_high": 0.6666666666666666
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.26417086507814297,
|
| 514 |
+
"f1_conclusion": 0.15384615384615385,
|
| 515 |
+
"f1_decree": 0.13333333333333333,
|
| 516 |
+
"f1_issue": 0.18181818181818182,
|
| 517 |
+
"f1_analysis": 0.5633802816901409,
|
| 518 |
+
"f1_facts": 0.13793103448275862,
|
| 519 |
+
"f1_procedural history": 0.3384615384615385,
|
| 520 |
+
"f1_rule": 0.3404255319148936,
|
| 521 |
+
"f1_macro_ci_low": 0.2096003414946908,
|
| 522 |
+
"f1_macro_ci_high": 0.3335047284451988,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.3076923076923077,
|
| 525 |
+
"score_ci_high": 0.37785141927318056,
|
| 526 |
+
"score_ci_low": 0.24264738787974433,
|
| 527 |
+
"num_of_instances": 200,
|
| 528 |
+
"accuracy": 0.25,
|
| 529 |
+
"accuracy_ci_low": 0.195,
|
| 530 |
+
"accuracy_ci_high": 0.315,
|
| 531 |
+
"f1_micro": 0.3076923076923077,
|
| 532 |
+
"f1_micro_ci_low": 0.24264738787974433,
|
| 533 |
+
"f1_micro_ci_high": 0.37785141927318056
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.463898065125059,
|
| 537 |
+
"f1_yes": 0.4662576687116564,
|
| 538 |
+
"f1_no": 0.46153846153846156,
|
| 539 |
+
"f1_macro_ci_low": 0.39561838155337925,
|
| 540 |
+
"f1_macro_ci_high": 0.5401290279855109,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.463855421686747,
|
| 543 |
+
"score_ci_high": 0.5388200914855517,
|
| 544 |
+
"score_ci_low": 0.3937587086040863,
|
| 545 |
+
"num_of_instances": 200,
|
| 546 |
+
"accuracy": 0.385,
|
| 547 |
+
"accuracy_ci_low": 0.32,
|
| 548 |
+
"accuracy_ci_high": 0.455,
|
| 549 |
+
"f1_micro": 0.463855421686747,
|
| 550 |
+
"f1_micro_ci_low": 0.3937587086040863,
|
| 551 |
+
"f1_micro_ci_high": 0.5388200914855517
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.75,
|
| 555 |
+
"f1_yes": 0.75,
|
| 556 |
+
"f1_no": 0.75,
|
| 557 |
+
"f1_macro_ci_low": 0.6597413488150221,
|
| 558 |
+
"f1_macro_ci_high": 0.8234621193211324,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.75,
|
| 561 |
+
"score_ci_high": 0.8234882632928148,
|
| 562 |
+
"score_ci_low": 0.6573326079878734,
|
| 563 |
+
"num_of_instances": 85,
|
| 564 |
+
"accuracy": 0.6352941176470588,
|
| 565 |
+
"accuracy_ci_low": 0.5411764705882353,
|
| 566 |
+
"accuracy_ci_high": 0.7294117647058823,
|
| 567 |
+
"f1_micro": 0.75,
|
| 568 |
+
"f1_micro_ci_low": 0.6573326079878734,
|
| 569 |
+
"f1_micro_ci_high": 0.8234882632928148
|
| 570 |
+
},
|
| 571 |
+
"score": 0.5552619268281919,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 770
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.6134758195135697,
|
| 578 |
+
"f1_cars": 0.8297872340425532,
|
| 579 |
+
"f1_windows x": 0.14492753623188406,
|
| 580 |
+
"f1_computer graphics": 0.41379310344827586,
|
| 581 |
+
"f1_atheism": 0.48148148148148145,
|
| 582 |
+
"f1_religion": 0.24561403508771928,
|
| 583 |
+
"f1_medicine": 0.8372093023255814,
|
| 584 |
+
"f1_christianity": 0.7878787878787878,
|
| 585 |
+
"f1_microsoft windows": 0.611764705882353,
|
| 586 |
+
"f1_middle east": 0.5970149253731343,
|
| 587 |
+
"f1_motorcycles": 0.7128712871287128,
|
| 588 |
+
"f1_pc hardware": 0.5736434108527132,
|
| 589 |
+
"f1_mac hardware": 0.6382978723404256,
|
| 590 |
+
"f1_electronics": 0.6436781609195402,
|
| 591 |
+
"f1_for sale": 0.7027027027027027,
|
| 592 |
+
"f1_guns": 0.34375,
|
| 593 |
+
"f1_space": 0.8269230769230769,
|
| 594 |
+
"f1_cryptography": 0.6486486486486487,
|
| 595 |
+
"f1_baseball": 0.9122807017543859,
|
| 596 |
+
"f1_politics": 0.3787878787878788,
|
| 597 |
+
"f1_hockey": 0.9384615384615385,
|
| 598 |
+
"f1_macro_ci_low": 0.5870232871618599,
|
| 599 |
+
"f1_macro_ci_high": 0.6399401092520309,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.6360655737704918,
|
| 602 |
+
"score_ci_high": 0.6652078896587189,
|
| 603 |
+
"score_ci_low": 0.606009312496833,
|
| 604 |
+
"num_of_instances": 1000,
|
| 605 |
+
"accuracy": 0.582,
|
| 606 |
+
"accuracy_ci_low": 0.549,
|
| 607 |
+
"accuracy_ci_high": 0.61,
|
| 608 |
+
"f1_micro": 0.6360655737704918,
|
| 609 |
+
"f1_micro_ci_low": 0.606009312496833,
|
| 610 |
+
"f1_micro_ci_high": 0.6652078896587189
|
| 611 |
+
},
|
| 612 |
+
"score": 0.6360655737704918,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 1000
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.7310250832696968,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.9301994301994302,
|
| 620 |
+
"f1_payday loan or title loan or personal loan": 0.2222222222222222,
|
| 621 |
+
"f1_mortgage": 0.8450704225352113,
|
| 622 |
+
"f1_credit card or prepaid card": 0.7596899224806202,
|
| 623 |
+
"f1_debt collection": 0.6887417218543046,
|
| 624 |
+
"f1_vehicle loan or lease": 0.7906976744186046,
|
| 625 |
+
"f1_checking or savings account": 0.8210526315789474,
|
| 626 |
+
"f1_money transfer or virtual currency or money service": 0.625,
|
| 627 |
+
"f1_student loan": 0.896551724137931,
|
| 628 |
+
"f1_macro_ci_low": 0.6859846742091775,
|
| 629 |
+
"f1_macro_ci_high": 0.7782039487360795,
|
| 630 |
+
"score_name": "f1_micro",
|
| 631 |
+
"score": 0.8752515090543259,
|
| 632 |
+
"score_ci_high": 0.8950697745652848,
|
| 633 |
+
"score_ci_low": 0.8535042875641251,
|
| 634 |
+
"num_of_instances": 1000,
|
| 635 |
+
"accuracy": 0.87,
|
| 636 |
+
"accuracy_ci_low": 0.848,
|
| 637 |
+
"accuracy_ci_high": 0.8900650893159325,
|
| 638 |
+
"f1_micro": 0.8752515090543259,
|
| 639 |
+
"f1_micro_ci_low": 0.8535042875641251,
|
| 640 |
+
"f1_micro_ci_high": 0.8950697745652848
|
| 641 |
+
},
|
| 642 |
+
"cfpb_product_watsonx": {
|
| 643 |
+
"f1_macro": 0.7446539429987555,
|
| 644 |
+
"f1_mortgages and loans": 0.8,
|
| 645 |
+
"f1_credit card": 0.7865168539325843,
|
| 646 |
+
"f1_debt collection": 0.6972477064220184,
|
| 647 |
+
"f1_credit reporting": 0.7835051546391752,
|
| 648 |
+
"f1_retail banking": 0.656,
|
| 649 |
+
"f1_macro_ci_low": 0.7076662872810797,
|
| 650 |
+
"f1_macro_ci_high": 0.7839689190392729,
|
| 651 |
+
"score_name": "f1_micro",
|
| 652 |
+
"score": 0.75177304964539,
|
| 653 |
+
"score_ci_high": 0.7885153719683515,
|
| 654 |
+
"score_ci_low": 0.7144890867826504,
|
| 655 |
+
"num_of_instances": 500,
|
| 656 |
+
"accuracy": 0.742,
|
| 657 |
+
"accuracy_ci_low": 0.7044918597766052,
|
| 658 |
+
"accuracy_ci_high": 0.78,
|
| 659 |
+
"f1_micro": 0.75177304964539,
|
| 660 |
+
"f1_micro_ci_low": 0.7144890867826504,
|
| 661 |
+
"f1_micro_ci_high": 0.7885153719683515
|
| 662 |
+
},
|
| 663 |
+
"score": 0.8135122793498579,
|
| 664 |
+
"score_name": "subsets_mean",
|
| 665 |
+
"num_of_instances": 1500
|
| 666 |
+
},
|
| 667 |
+
"qa_finance": {
|
| 668 |
+
"fin_qa": {
|
| 669 |
+
"num_of_instances": 1000,
|
| 670 |
+
"program_accuracy": 0.206,
|
| 671 |
+
"score": 0.206,
|
| 672 |
+
"score_name": "program_accuracy",
|
| 673 |
+
"execution_accuracy": 0.183,
|
| 674 |
+
"program_accuracy_ci_low": 0.18192581131748398,
|
| 675 |
+
"program_accuracy_ci_high": 0.233,
|
| 676 |
+
"score_ci_low": 0.18192581131748398,
|
| 677 |
+
"score_ci_high": 0.233,
|
| 678 |
+
"execution_accuracy_ci_low": 0.16,
|
| 679 |
+
"execution_accuracy_ci_high": 0.206
|
| 680 |
+
},
|
| 681 |
+
"score": 0.206,
|
| 682 |
+
"score_name": "subsets_mean",
|
| 683 |
+
"num_of_instances": 1000
|
| 684 |
+
},
|
| 685 |
+
"rag_general": {
|
| 686 |
+
"rag_response_generation_clapnq": {
|
| 687 |
+
"precision": 0.2829195176672266,
|
| 688 |
+
"recall": 0.5895069651787405,
|
| 689 |
+
"f1": 0.3247642449571121,
|
| 690 |
+
"precision_ci_low": 0.2632314020521836,
|
| 691 |
+
"precision_ci_high": 0.3018747287495355,
|
| 692 |
+
"recall_ci_low": 0.5757048544050242,
|
| 693 |
+
"recall_ci_high": 0.605461197768975,
|
| 694 |
+
"f1_ci_low": 0.3070149800069332,
|
| 695 |
+
"f1_ci_high": 0.3415261837814938,
|
| 696 |
+
"score_name": "f1",
|
| 697 |
+
"score": 0.3247642449571121,
|
| 698 |
+
"score_ci_high": 0.3415261837814938,
|
| 699 |
+
"score_ci_low": 0.3070149800069332,
|
| 700 |
+
"num_of_instances": 600,
|
| 701 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.5840552796920141,
|
| 702 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.6764627607663473,
|
| 703 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.5322982901086409,
|
| 704 |
+
"faithfullness_f1_token_overlap": 0.3161988054051969,
|
| 705 |
+
"faithfullness_recall_token_overlap": 0.2527973404819663,
|
| 706 |
+
"faithfullness_precision_token_overlap": 0.5297932448406707,
|
| 707 |
+
"correctness_f1_token_overlap": 0.3247642449571121,
|
| 708 |
+
"correctness_recall_token_overlap": 0.5895069651787405,
|
| 709 |
+
"correctness_precision_token_overlap": 0.2829195176672266
|
| 710 |
+
},
|
| 711 |
+
"score": 0.3247642449571121,
|
| 712 |
+
"score_name": "subsets_mean",
|
| 713 |
+
"num_of_instances": 600
|
| 714 |
+
},
|
| 715 |
+
"reasoning": {
|
| 716 |
+
"hellaswag": {
|
| 717 |
+
"accuracy": 0.568,
|
| 718 |
+
"accuracy_ci_low": 0.536,
|
| 719 |
+
"accuracy_ci_high": 0.597,
|
| 720 |
+
"score_name": "accuracy",
|
| 721 |
+
"score": 0.568,
|
| 722 |
+
"score_ci_high": 0.597,
|
| 723 |
+
"score_ci_low": 0.536,
|
| 724 |
+
"num_of_instances": 1000
|
| 725 |
+
},
|
| 726 |
+
"openbook_qa": {
|
| 727 |
+
"accuracy": 0.922,
|
| 728 |
+
"accuracy_ci_low": 0.8937961961124387,
|
| 729 |
+
"accuracy_ci_high": 0.9425849640401368,
|
| 730 |
+
"score_name": "accuracy",
|
| 731 |
+
"score": 0.922,
|
| 732 |
+
"score_ci_high": 0.9425849640401368,
|
| 733 |
+
"score_ci_low": 0.8937961961124387,
|
| 734 |
+
"num_of_instances": 500
|
| 735 |
+
},
|
| 736 |
+
"score": 0.745,
|
| 737 |
+
"score_name": "subsets_mean",
|
| 738 |
+
"num_of_instances": 1500
|
| 739 |
+
},
|
| 740 |
+
"safety": {
|
| 741 |
+
"attaq_500": {
|
| 742 |
+
"severity_critical": 0.0,
|
| 743 |
+
"severity_high": 0.0,
|
| 744 |
+
"severity_medium": 1.0,
|
| 745 |
+
"severity_low": 99.0,
|
| 746 |
+
"category_harmful_info": 0.8334540272453089,
|
| 747 |
+
"category_deception": 0.8512123076602666,
|
| 748 |
+
"category_explicit_content": 0.8684844600431846,
|
| 749 |
+
"category_violence": 0.8207998569862591,
|
| 750 |
+
"category_discrimination": 0.7722197345324925,
|
| 751 |
+
"category_substance_abuse": 0.8600103222661548,
|
| 752 |
+
"category_pii": 0.8025329068855003,
|
| 753 |
+
"safety": 0.8309456067584041,
|
| 754 |
+
"safety_ci_low": 0.812541088262446,
|
| 755 |
+
"safety_ci_high": 0.8500828095785149,
|
| 756 |
+
"score_name": "safety",
|
| 757 |
+
"score": 0.8309456067584041,
|
| 758 |
+
"score_ci_high": 0.8500828095785149,
|
| 759 |
+
"score_ci_low": 0.812541088262446,
|
| 760 |
+
"num_of_instances": 100
|
| 761 |
+
},
|
| 762 |
+
"score": 0.8309456067584041,
|
| 763 |
+
"score_name": "subsets_mean",
|
| 764 |
+
"num_of_instances": 100
|
| 765 |
+
},
|
| 766 |
+
"summarization": {
|
| 767 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 768 |
+
"num_of_instances": 528,
|
| 769 |
+
"rouge1": 0.4363286082688575,
|
| 770 |
+
"rouge2": 0.22072608165229177,
|
| 771 |
+
"rougeL": 0.3067201488969288,
|
| 772 |
+
"score": 0.3067201488969288,
|
| 773 |
+
"score_name": "rougeL",
|
| 774 |
+
"rougeLsum": 0.37348424474826925,
|
| 775 |
+
"rouge1_ci_low": 0.42622540172251744,
|
| 776 |
+
"rouge1_ci_high": 0.445479768051287,
|
| 777 |
+
"rouge2_ci_low": 0.2133469390728694,
|
| 778 |
+
"rouge2_ci_high": 0.22852486335907615,
|
| 779 |
+
"rougeL_ci_low": 0.2990472341490495,
|
| 780 |
+
"rougeL_ci_high": 0.31469167779345575,
|
| 781 |
+
"score_ci_low": 0.2990472341490495,
|
| 782 |
+
"score_ci_high": 0.31469167779345575,
|
| 783 |
+
"rougeLsum_ci_low": 0.36439724635298765,
|
| 784 |
+
"rougeLsum_ci_high": 0.3821672398567385
|
| 785 |
+
},
|
| 786 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 787 |
+
"num_of_instances": 1000,
|
| 788 |
+
"rouge1": 0.13281253260941045,
|
| 789 |
+
"rouge2": 0.020216836126571446,
|
| 790 |
+
"rougeL": 0.0953008337162679,
|
| 791 |
+
"score": 0.0953008337162679,
|
| 792 |
+
"score_name": "rougeL",
|
| 793 |
+
"rougeLsum": 0.10871871026938495,
|
| 794 |
+
"rouge1_ci_low": 0.1270654167753335,
|
| 795 |
+
"rouge1_ci_high": 0.13798234251944885,
|
| 796 |
+
"rouge2_ci_low": 0.01830191124480353,
|
| 797 |
+
"rouge2_ci_high": 0.022113898574922495,
|
| 798 |
+
"rougeL_ci_low": 0.09124510500356316,
|
| 799 |
+
"rougeL_ci_high": 0.098681444438937,
|
| 800 |
+
"score_ci_low": 0.09124510500356316,
|
| 801 |
+
"score_ci_high": 0.098681444438937,
|
| 802 |
+
"rougeLsum_ci_low": 0.10412203835097511,
|
| 803 |
+
"rougeLsum_ci_high": 0.11266943648159043
|
| 804 |
+
},
|
| 805 |
+
"score": 0.20101049130659834,
|
| 806 |
+
"score_name": "subsets_mean",
|
| 807 |
+
"num_of_instances": 1528
|
| 808 |
+
},
|
| 809 |
+
"translation": {
|
| 810 |
+
"mt_flores_101_ara_eng": {
|
| 811 |
+
"num_of_instances": 66,
|
| 812 |
+
"counts": [
|
| 813 |
+
1297,
|
| 814 |
+
853,
|
| 815 |
+
602,
|
| 816 |
+
432
|
| 817 |
+
],
|
| 818 |
+
"totals": [
|
| 819 |
+
1773,
|
| 820 |
+
1707,
|
| 821 |
+
1641,
|
| 822 |
+
1575
|
| 823 |
+
],
|
| 824 |
+
"precisions": [
|
| 825 |
+
0.7315284827975184,
|
| 826 |
+
0.4997070884592853,
|
| 827 |
+
0.3668494820231566,
|
| 828 |
+
0.27428571428571424
|
| 829 |
+
],
|
| 830 |
+
"bp": 1.0,
|
| 831 |
+
"sys_len": 1773,
|
| 832 |
+
"ref_len": 1734,
|
| 833 |
+
"sacrebleu": 0.4379348051104114,
|
| 834 |
+
"score": 0.4379348051104114,
|
| 835 |
+
"score_name": "sacrebleu",
|
| 836 |
+
"score_ci_low": 0.3839572779473718,
|
| 837 |
+
"score_ci_high": 0.47542268740294724,
|
| 838 |
+
"sacrebleu_ci_low": 0.3839572779473718,
|
| 839 |
+
"sacrebleu_ci_high": 0.47542268740294724
|
| 840 |
+
},
|
| 841 |
+
"mt_flores_101_deu_eng": {
|
| 842 |
+
"num_of_instances": 66,
|
| 843 |
+
"counts": [
|
| 844 |
+
1336,
|
| 845 |
+
889,
|
| 846 |
+
624,
|
| 847 |
+
446
|
| 848 |
+
],
|
| 849 |
+
"totals": [
|
| 850 |
+
1793,
|
| 851 |
+
1727,
|
| 852 |
+
1661,
|
| 853 |
+
1595
|
| 854 |
+
],
|
| 855 |
+
"precisions": [
|
| 856 |
+
0.7451199107640825,
|
| 857 |
+
0.5147654892877823,
|
| 858 |
+
0.37567730282962075,
|
| 859 |
+
0.27962382445141065
|
| 860 |
+
],
|
| 861 |
+
"bp": 1.0,
|
| 862 |
+
"sys_len": 1793,
|
| 863 |
+
"ref_len": 1734,
|
| 864 |
+
"sacrebleu": 0.4480290559468694,
|
| 865 |
+
"score": 0.4480290559468694,
|
| 866 |
+
"score_name": "sacrebleu",
|
| 867 |
+
"score_ci_low": 0.40758551741171895,
|
| 868 |
+
"score_ci_high": 0.49424014474329564,
|
| 869 |
+
"sacrebleu_ci_low": 0.40758551741171895,
|
| 870 |
+
"sacrebleu_ci_high": 0.49424014474329564
|
| 871 |
+
},
|
| 872 |
+
"mt_flores_101_eng_ara": {
|
| 873 |
+
"num_of_instances": 66,
|
| 874 |
+
"counts": [
|
| 875 |
+
908,
|
| 876 |
+
498,
|
| 877 |
+
301,
|
| 878 |
+
181
|
| 879 |
+
],
|
| 880 |
+
"totals": [
|
| 881 |
+
1600,
|
| 882 |
+
1534,
|
| 883 |
+
1468,
|
| 884 |
+
1402
|
| 885 |
+
],
|
| 886 |
+
"precisions": [
|
| 887 |
+
0.5675,
|
| 888 |
+
0.3246414602346806,
|
| 889 |
+
0.20504087193460488,
|
| 890 |
+
0.1291012838801712
|
| 891 |
+
],
|
| 892 |
+
"bp": 1.0,
|
| 893 |
+
"sys_len": 1600,
|
| 894 |
+
"ref_len": 1589,
|
| 895 |
+
"sacrebleu": 0.26426230123240096,
|
| 896 |
+
"score": 0.26426230123240096,
|
| 897 |
+
"score_name": "sacrebleu",
|
| 898 |
+
"score_ci_low": 0.2206088677339497,
|
| 899 |
+
"score_ci_high": 0.3043911445742578,
|
| 900 |
+
"sacrebleu_ci_low": 0.2206088677339497,
|
| 901 |
+
"sacrebleu_ci_high": 0.3043911445742578
|
| 902 |
+
},
|
| 903 |
+
"mt_flores_101_eng_deu": {
|
| 904 |
+
"num_of_instances": 66,
|
| 905 |
+
"counts": [
|
| 906 |
+
1252,
|
| 907 |
+
762,
|
| 908 |
+
513,
|
| 909 |
+
371
|
| 910 |
+
],
|
| 911 |
+
"totals": [
|
| 912 |
+
1850,
|
| 913 |
+
1784,
|
| 914 |
+
1718,
|
| 915 |
+
1652
|
| 916 |
+
],
|
| 917 |
+
"precisions": [
|
| 918 |
+
0.6767567567567567,
|
| 919 |
+
0.42713004484304934,
|
| 920 |
+
0.29860302677532014,
|
| 921 |
+
0.2245762711864407
|
| 922 |
+
],
|
| 923 |
+
"bp": 1.0,
|
| 924 |
+
"sys_len": 1850,
|
| 925 |
+
"ref_len": 1835,
|
| 926 |
+
"sacrebleu": 0.37313217401540816,
|
| 927 |
+
"score": 0.37313217401540816,
|
| 928 |
+
"score_name": "sacrebleu",
|
| 929 |
+
"score_ci_low": 0.3292059035458714,
|
| 930 |
+
"score_ci_high": 0.4247888842695035,
|
| 931 |
+
"sacrebleu_ci_low": 0.3292059035458714,
|
| 932 |
+
"sacrebleu_ci_high": 0.4247888842695035
|
| 933 |
+
},
|
| 934 |
+
"mt_flores_101_eng_fra": {
|
| 935 |
+
"num_of_instances": 66,
|
| 936 |
+
"counts": [
|
| 937 |
+
1553,
|
| 938 |
+
1160,
|
| 939 |
+
918,
|
| 940 |
+
734
|
| 941 |
+
],
|
| 942 |
+
"totals": [
|
| 943 |
+
2035,
|
| 944 |
+
1969,
|
| 945 |
+
1903,
|
| 946 |
+
1837
|
| 947 |
+
],
|
| 948 |
+
"precisions": [
|
| 949 |
+
0.763144963144963,
|
| 950 |
+
0.5891315388522093,
|
| 951 |
+
0.4823962165002627,
|
| 952 |
+
0.3995645073489385
|
| 953 |
+
],
|
| 954 |
+
"bp": 0.9839145587719164,
|
| 955 |
+
"sys_len": 2035,
|
| 956 |
+
"ref_len": 2068,
|
| 957 |
+
"sacrebleu": 0.5338385284906572,
|
| 958 |
+
"score": 0.5338385284906572,
|
| 959 |
+
"score_name": "sacrebleu",
|
| 960 |
+
"score_ci_low": 0.4960508343377122,
|
| 961 |
+
"score_ci_high": 0.5778992367091751,
|
| 962 |
+
"sacrebleu_ci_low": 0.4960508343377122,
|
| 963 |
+
"sacrebleu_ci_high": 0.5778992367091751
|
| 964 |
+
},
|
| 965 |
+
"mt_flores_101_eng_kor": {
|
| 966 |
+
"num_of_instances": 66,
|
| 967 |
+
"counts": [
|
| 968 |
+
1322,
|
| 969 |
+
700,
|
| 970 |
+
421,
|
| 971 |
+
259
|
| 972 |
+
],
|
| 973 |
+
"totals": [
|
| 974 |
+
2388,
|
| 975 |
+
2322,
|
| 976 |
+
2256,
|
| 977 |
+
2190
|
| 978 |
+
],
|
| 979 |
+
"precisions": [
|
| 980 |
+
0.5536013400335008,
|
| 981 |
+
0.301464254952627,
|
| 982 |
+
0.18661347517730498,
|
| 983 |
+
0.1182648401826484
|
| 984 |
+
],
|
| 985 |
+
"bp": 1.0,
|
| 986 |
+
"sys_len": 2388,
|
| 987 |
+
"ref_len": 2235,
|
| 988 |
+
"sacrebleu": 0.2463530263425343,
|
| 989 |
+
"score": 0.2463530263425343,
|
| 990 |
+
"score_name": "sacrebleu",
|
| 991 |
+
"score_ci_low": 0.21902218576525906,
|
| 992 |
+
"score_ci_high": 0.28268262467368716,
|
| 993 |
+
"sacrebleu_ci_low": 0.21902218576525906,
|
| 994 |
+
"sacrebleu_ci_high": 0.28268262467368716
|
| 995 |
+
},
|
| 996 |
+
"mt_flores_101_eng_por": {
|
| 997 |
+
"num_of_instances": 66,
|
| 998 |
+
"counts": [
|
| 999 |
+
1463,
|
| 1000 |
+
1077,
|
| 1001 |
+
833,
|
| 1002 |
+
645
|
| 1003 |
+
],
|
| 1004 |
+
"totals": [
|
| 1005 |
+
1890,
|
| 1006 |
+
1824,
|
| 1007 |
+
1758,
|
| 1008 |
+
1692
|
| 1009 |
+
],
|
| 1010 |
+
"precisions": [
|
| 1011 |
+
0.774074074074074,
|
| 1012 |
+
0.5904605263157895,
|
| 1013 |
+
0.47383390216154725,
|
| 1014 |
+
0.3812056737588652
|
| 1015 |
+
],
|
| 1016 |
+
"bp": 0.9863375760488048,
|
| 1017 |
+
"sys_len": 1890,
|
| 1018 |
+
"ref_len": 1916,
|
| 1019 |
+
"sacrebleu": 0.5287072892340219,
|
| 1020 |
+
"score": 0.5287072892340219,
|
| 1021 |
+
"score_name": "sacrebleu",
|
| 1022 |
+
"score_ci_low": 0.47933696238952217,
|
| 1023 |
+
"score_ci_high": 0.5714963999565859,
|
| 1024 |
+
"sacrebleu_ci_low": 0.47933696238952217,
|
| 1025 |
+
"sacrebleu_ci_high": 0.5714963999565859
|
| 1026 |
+
},
|
| 1027 |
+
"mt_flores_101_eng_ron": {
|
| 1028 |
+
"num_of_instances": 66,
|
| 1029 |
+
"counts": [
|
| 1030 |
+
1402,
|
| 1031 |
+
981,
|
| 1032 |
+
716,
|
| 1033 |
+
523
|
| 1034 |
+
],
|
| 1035 |
+
"totals": [
|
| 1036 |
+
1969,
|
| 1037 |
+
1903,
|
| 1038 |
+
1837,
|
| 1039 |
+
1771
|
| 1040 |
+
],
|
| 1041 |
+
"precisions": [
|
| 1042 |
+
0.71203656678517,
|
| 1043 |
+
0.5155018392012611,
|
| 1044 |
+
0.3897659227000544,
|
| 1045 |
+
0.295313382269904
|
| 1046 |
+
],
|
| 1047 |
+
"bp": 1.0,
|
| 1048 |
+
"sys_len": 1969,
|
| 1049 |
+
"ref_len": 1949,
|
| 1050 |
+
"sacrebleu": 0.4533721907576224,
|
| 1051 |
+
"score": 0.4533721907576224,
|
| 1052 |
+
"score_name": "sacrebleu",
|
| 1053 |
+
"score_ci_low": 0.41581094731566015,
|
| 1054 |
+
"score_ci_high": 0.498033544687375,
|
| 1055 |
+
"sacrebleu_ci_low": 0.41581094731566015,
|
| 1056 |
+
"sacrebleu_ci_high": 0.498033544687375
|
| 1057 |
+
},
|
| 1058 |
+
"mt_flores_101_eng_spa": {
|
| 1059 |
+
"num_of_instances": 66,
|
| 1060 |
+
"counts": [
|
| 1061 |
+
1285,
|
| 1062 |
+
735,
|
| 1063 |
+
453,
|
| 1064 |
+
283
|
| 1065 |
+
],
|
| 1066 |
+
"totals": [
|
| 1067 |
+
1994,
|
| 1068 |
+
1928,
|
| 1069 |
+
1862,
|
| 1070 |
+
1796
|
| 1071 |
+
],
|
| 1072 |
+
"precisions": [
|
| 1073 |
+
0.6444332998996991,
|
| 1074 |
+
0.38122406639004147,
|
| 1075 |
+
0.24328678839957035,
|
| 1076 |
+
0.15757238307349666
|
| 1077 |
+
],
|
| 1078 |
+
"bp": 0.9491803375373334,
|
| 1079 |
+
"sys_len": 1994,
|
| 1080 |
+
"ref_len": 2098,
|
| 1081 |
+
"sacrebleu": 0.2956909013134059,
|
| 1082 |
+
"score": 0.2956909013134059,
|
| 1083 |
+
"score_name": "sacrebleu",
|
| 1084 |
+
"score_ci_low": 0.2691627285024834,
|
| 1085 |
+
"score_ci_high": 0.32834717984418194,
|
| 1086 |
+
"sacrebleu_ci_low": 0.2691627285024834,
|
| 1087 |
+
"sacrebleu_ci_high": 0.32834717984418194
|
| 1088 |
+
},
|
| 1089 |
+
"mt_flores_101_fra_eng": {
|
| 1090 |
+
"num_of_instances": 66,
|
| 1091 |
+
"counts": [
|
| 1092 |
+
1366,
|
| 1093 |
+
964,
|
| 1094 |
+
696,
|
| 1095 |
+
499
|
| 1096 |
+
],
|
| 1097 |
+
"totals": [
|
| 1098 |
+
1844,
|
| 1099 |
+
1778,
|
| 1100 |
+
1712,
|
| 1101 |
+
1646
|
| 1102 |
+
],
|
| 1103 |
+
"precisions": [
|
| 1104 |
+
0.7407809110629067,
|
| 1105 |
+
0.5421822272215973,
|
| 1106 |
+
0.40654205607476634,
|
| 1107 |
+
0.30315917375455653
|
| 1108 |
+
],
|
| 1109 |
+
"bp": 1.0,
|
| 1110 |
+
"sys_len": 1844,
|
| 1111 |
+
"ref_len": 1734,
|
| 1112 |
+
"sacrebleu": 0.47168581235777585,
|
| 1113 |
+
"score": 0.47168581235777585,
|
| 1114 |
+
"score_name": "sacrebleu",
|
| 1115 |
+
"score_ci_low": 0.41832464742975967,
|
| 1116 |
+
"score_ci_high": 0.5113720236775223,
|
| 1117 |
+
"sacrebleu_ci_low": 0.41832464742975967,
|
| 1118 |
+
"sacrebleu_ci_high": 0.5113720236775223
|
| 1119 |
+
},
|
| 1120 |
+
"mt_flores_101_jpn_eng": {
|
| 1121 |
+
"num_of_instances": 66,
|
| 1122 |
+
"counts": [
|
| 1123 |
+
1121,
|
| 1124 |
+
627,
|
| 1125 |
+
398,
|
| 1126 |
+
260
|
| 1127 |
+
],
|
| 1128 |
+
"totals": [
|
| 1129 |
+
1831,
|
| 1130 |
+
1765,
|
| 1131 |
+
1699,
|
| 1132 |
+
1633
|
| 1133 |
+
],
|
| 1134 |
+
"precisions": [
|
| 1135 |
+
0.6122337520480612,
|
| 1136 |
+
0.35524079320113316,
|
| 1137 |
+
0.23425544437904647,
|
| 1138 |
+
0.15921616656460502
|
| 1139 |
+
],
|
| 1140 |
+
"bp": 1.0,
|
| 1141 |
+
"sys_len": 1831,
|
| 1142 |
+
"ref_len": 1734,
|
| 1143 |
+
"sacrebleu": 0.30010915149877077,
|
| 1144 |
+
"score": 0.30010915149877077,
|
| 1145 |
+
"score_name": "sacrebleu",
|
| 1146 |
+
"score_ci_low": 0.25220636468972635,
|
| 1147 |
+
"score_ci_high": 0.34253906716484656,
|
| 1148 |
+
"sacrebleu_ci_low": 0.25220636468972635,
|
| 1149 |
+
"sacrebleu_ci_high": 0.34253906716484656
|
| 1150 |
+
},
|
| 1151 |
+
"mt_flores_101_kor_eng": {
|
| 1152 |
+
"num_of_instances": 66,
|
| 1153 |
+
"counts": [
|
| 1154 |
+
1108,
|
| 1155 |
+
607,
|
| 1156 |
+
365,
|
| 1157 |
+
233
|
| 1158 |
+
],
|
| 1159 |
+
"totals": [
|
| 1160 |
+
1751,
|
| 1161 |
+
1685,
|
| 1162 |
+
1619,
|
| 1163 |
+
1553
|
| 1164 |
+
],
|
| 1165 |
+
"precisions": [
|
| 1166 |
+
0.6327812678469446,
|
| 1167 |
+
0.3602373887240356,
|
| 1168 |
+
0.22544780728844965,
|
| 1169 |
+
0.150032195750161
|
| 1170 |
+
],
|
| 1171 |
+
"bp": 1.0,
|
| 1172 |
+
"sys_len": 1751,
|
| 1173 |
+
"ref_len": 1734,
|
| 1174 |
+
"sacrebleu": 0.2963249503420515,
|
| 1175 |
+
"score": 0.2963249503420515,
|
| 1176 |
+
"score_name": "sacrebleu",
|
| 1177 |
+
"score_ci_low": 0.25949260863076823,
|
| 1178 |
+
"score_ci_high": 0.3478140976036152,
|
| 1179 |
+
"sacrebleu_ci_low": 0.25949260863076823,
|
| 1180 |
+
"sacrebleu_ci_high": 0.3478140976036152
|
| 1181 |
+
},
|
| 1182 |
+
"mt_flores_101_por_eng": {
|
| 1183 |
+
"num_of_instances": 66,
|
| 1184 |
+
"counts": [
|
| 1185 |
+
1380,
|
| 1186 |
+
1014,
|
| 1187 |
+
776,
|
| 1188 |
+
594
|
| 1189 |
+
],
|
| 1190 |
+
"totals": [
|
| 1191 |
+
1815,
|
| 1192 |
+
1749,
|
| 1193 |
+
1683,
|
| 1194 |
+
1617
|
| 1195 |
+
],
|
| 1196 |
+
"precisions": [
|
| 1197 |
+
0.7603305785123967,
|
| 1198 |
+
0.5797598627787307,
|
| 1199 |
+
0.46108140225787286,
|
| 1200 |
+
0.3673469387755102
|
| 1201 |
+
],
|
| 1202 |
+
"bp": 1.0,
|
| 1203 |
+
"sys_len": 1815,
|
| 1204 |
+
"ref_len": 1734,
|
| 1205 |
+
"sacrebleu": 0.5227284747817351,
|
| 1206 |
+
"score": 0.5227284747817351,
|
| 1207 |
+
"score_name": "sacrebleu",
|
| 1208 |
+
"score_ci_low": 0.46746182790145246,
|
| 1209 |
+
"score_ci_high": 0.5686400938537622,
|
| 1210 |
+
"sacrebleu_ci_low": 0.46746182790145246,
|
| 1211 |
+
"sacrebleu_ci_high": 0.5686400938537622
|
| 1212 |
+
},
|
| 1213 |
+
"mt_flores_101_ron_eng": {
|
| 1214 |
+
"num_of_instances": 66,
|
| 1215 |
+
"counts": [
|
| 1216 |
+
1375,
|
| 1217 |
+
972,
|
| 1218 |
+
711,
|
| 1219 |
+
529
|
| 1220 |
+
],
|
| 1221 |
+
"totals": [
|
| 1222 |
+
1800,
|
| 1223 |
+
1734,
|
| 1224 |
+
1668,
|
| 1225 |
+
1602
|
| 1226 |
+
],
|
| 1227 |
+
"precisions": [
|
| 1228 |
+
0.7638888888888888,
|
| 1229 |
+
0.560553633217993,
|
| 1230 |
+
0.4262589928057554,
|
| 1231 |
+
0.33021223470661676
|
| 1232 |
+
],
|
| 1233 |
+
"bp": 1.0,
|
| 1234 |
+
"sys_len": 1800,
|
| 1235 |
+
"ref_len": 1734,
|
| 1236 |
+
"sacrebleu": 0.49548272782749336,
|
| 1237 |
+
"score": 0.49548272782749336,
|
| 1238 |
+
"score_name": "sacrebleu",
|
| 1239 |
+
"score_ci_low": 0.45134581590558903,
|
| 1240 |
+
"score_ci_high": 0.5288737168538972,
|
| 1241 |
+
"sacrebleu_ci_low": 0.45134581590558903,
|
| 1242 |
+
"sacrebleu_ci_high": 0.5288737168538972
|
| 1243 |
+
},
|
| 1244 |
+
"mt_flores_101_spa_eng": {
|
| 1245 |
+
"num_of_instances": 66,
|
| 1246 |
+
"counts": [
|
| 1247 |
+
1218,
|
| 1248 |
+
705,
|
| 1249 |
+
454,
|
| 1250 |
+
300
|
| 1251 |
+
],
|
| 1252 |
+
"totals": [
|
| 1253 |
+
1885,
|
| 1254 |
+
1819,
|
| 1255 |
+
1753,
|
| 1256 |
+
1687
|
| 1257 |
+
],
|
| 1258 |
+
"precisions": [
|
| 1259 |
+
0.6461538461538461,
|
| 1260 |
+
0.38757559098405714,
|
| 1261 |
+
0.2589845978322875,
|
| 1262 |
+
0.17783046828689983
|
| 1263 |
+
],
|
| 1264 |
+
"bp": 1.0,
|
| 1265 |
+
"sys_len": 1885,
|
| 1266 |
+
"ref_len": 1734,
|
| 1267 |
+
"sacrebleu": 0.32771258154747107,
|
| 1268 |
+
"score": 0.32771258154747107,
|
| 1269 |
+
"score_name": "sacrebleu",
|
| 1270 |
+
"score_ci_low": 0.3009156083594827,
|
| 1271 |
+
"score_ci_high": 0.3820347897282282,
|
| 1272 |
+
"sacrebleu_ci_low": 0.3009156083594827,
|
| 1273 |
+
"sacrebleu_ci_high": 0.3820347897282282
|
| 1274 |
+
},
|
| 1275 |
+
"score": 0.3996909313865753,
|
| 1276 |
+
"score_name": "subsets_mean",
|
| 1277 |
+
"num_of_instances": 990
|
| 1278 |
+
},
|
| 1279 |
+
"score": 0.5400687869936714,
|
| 1280 |
+
"score_name": "subsets_mean",
|
| 1281 |
+
"num_of_instances": 12472
|
| 1282 |
+
}
|
| 1283 |
+
}
|
results/bluebench/2025-06-21T11-34-24_evaluation_results.json
ADDED
|
@@ -0,0 +1,1283 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"environment_info": {
|
| 3 |
+
"timestamp_utc": "2025-06-21T15:34:20.869271Z",
|
| 4 |
+
"command_line_invocation": [
|
| 5 |
+
"/dccstor/jbworks/miniforge3/envs/bb/bin/unitxt-evaluate",
|
| 6 |
+
"--tasks",
|
| 7 |
+
"benchmarks.bluebench",
|
| 8 |
+
"--model",
|
| 9 |
+
"cross_provider",
|
| 10 |
+
"--model_args",
|
| 11 |
+
"model_name=watsonx/mistralai/mistral-small-3-1-24b-instruct-2503,max_tokens=256",
|
| 12 |
+
"--output_path",
|
| 13 |
+
"./results/bluebench",
|
| 14 |
+
"--log_samples",
|
| 15 |
+
"--trust_remote_code",
|
| 16 |
+
"--batch_size",
|
| 17 |
+
"8",
|
| 18 |
+
"--verbosity",
|
| 19 |
+
"ERROR"
|
| 20 |
+
],
|
| 21 |
+
"parsed_arguments": {
|
| 22 |
+
"tasks": [
|
| 23 |
+
"benchmarks.bluebench"
|
| 24 |
+
],
|
| 25 |
+
"split": "test",
|
| 26 |
+
"num_fewshots": null,
|
| 27 |
+
"limit": null,
|
| 28 |
+
"batch_size": 8,
|
| 29 |
+
"model": "watsonx/mistralai/mistral-small-3-1-24b-instruct-2503",
|
| 30 |
+
"model_args": {
|
| 31 |
+
"max_tokens": 256
|
| 32 |
+
},
|
| 33 |
+
"gen_kwargs": null,
|
| 34 |
+
"chat_template_kwargs": null,
|
| 35 |
+
"output_path": "./results/bluebench",
|
| 36 |
+
"output_file_prefix": "evaluation_results",
|
| 37 |
+
"log_samples": true,
|
| 38 |
+
"verbosity": "ERROR",
|
| 39 |
+
"apply_chat_template": false,
|
| 40 |
+
"trust_remote_code": true,
|
| 41 |
+
"disable_hf_cache": false,
|
| 42 |
+
"cache_dir": null
|
| 43 |
+
},
|
| 44 |
+
"unitxt_version": "1.24.0",
|
| 45 |
+
"unitxt_commit_hash": "2bfd4494ec443ef86013e30d31f4860177124476",
|
| 46 |
+
"python_version": "3.10.18",
|
| 47 |
+
"system": "Linux",
|
| 48 |
+
"system_version": "#1 SMP PREEMPT_DYNAMIC Fri Aug 9 14:06:03 EDT 2024",
|
| 49 |
+
"installed_packages": {
|
| 50 |
+
"nvidia-cufile-cu12": "1.11.1.6",
|
| 51 |
+
"triton": "3.3.1",
|
| 52 |
+
"nltk": "3.9.1",
|
| 53 |
+
"anyio": "4.9.0",
|
| 54 |
+
"absl-py": "2.3.0",
|
| 55 |
+
"tiktoken": "0.9.0",
|
| 56 |
+
"charset-normalizer": "3.4.2",
|
| 57 |
+
"nvidia-cuda-runtime-cu12": "12.6.77",
|
| 58 |
+
"sympy": "1.14.0",
|
| 59 |
+
"mecab-ko": "1.0.1",
|
| 60 |
+
"litellm": "1.72.6.post1",
|
| 61 |
+
"httpcore": "1.0.9",
|
| 62 |
+
"Jinja2": "3.1.6",
|
| 63 |
+
"jsonschema-specifications": "2025.4.1",
|
| 64 |
+
"pydantic_core": "2.33.2",
|
| 65 |
+
"nvidia-cusparse-cu12": "12.5.4.2",
|
| 66 |
+
"yarl": "1.20.1",
|
| 67 |
+
"openai": "1.88.0",
|
| 68 |
+
"portalocker": "3.2.0",
|
| 69 |
+
"pandas": "2.3.0",
|
| 70 |
+
"multiprocess": "0.70.16",
|
| 71 |
+
"jsonschema": "4.24.0",
|
| 72 |
+
"unitxt": "1.24.0",
|
| 73 |
+
"nvidia-nvjitlink-cu12": "12.6.85",
|
| 74 |
+
"nvidia-cublas-cu12": "12.6.4.1",
|
| 75 |
+
"pydantic": "2.11.7",
|
| 76 |
+
"async-timeout": "5.0.1",
|
| 77 |
+
"annotated-types": "0.7.0",
|
| 78 |
+
"rouge_score": "0.1.2",
|
| 79 |
+
"contourpy": "1.3.2",
|
| 80 |
+
"aiosignal": "1.3.2",
|
| 81 |
+
"nvidia-cuda-cupti-cu12": "12.6.80",
|
| 82 |
+
"pillow": "11.2.1",
|
| 83 |
+
"six": "1.17.0",
|
| 84 |
+
"diskcache": "5.6.3",
|
| 85 |
+
"tqdm": "4.67.1",
|
| 86 |
+
"pyarrow": "20.0.0",
|
| 87 |
+
"h11": "0.16.0",
|
| 88 |
+
"zipp": "3.19.2",
|
| 89 |
+
"tzdata": "2025.2",
|
| 90 |
+
"bert-score": "0.3.13",
|
| 91 |
+
"setuptools": "80.9.0",
|
| 92 |
+
"referencing": "0.36.2",
|
| 93 |
+
"sacrebleu": "2.5.1",
|
| 94 |
+
"filelock": "3.18.0",
|
| 95 |
+
"urllib3": "2.5.0",
|
| 96 |
+
"scipy": "1.15.3",
|
| 97 |
+
"nvidia-nccl-cu12": "2.26.2",
|
| 98 |
+
"kiwisolver": "1.4.8",
|
| 99 |
+
"networkx": "3.4.2",
|
| 100 |
+
"typing-inspection": "0.4.1",
|
| 101 |
+
"lxml": "5.4.0",
|
| 102 |
+
"sniffio": "1.3.1",
|
| 103 |
+
"scikit-learn": "1.7.0",
|
| 104 |
+
"nvidia-curand-cu12": "10.3.7.77",
|
| 105 |
+
"pip": "25.1.1",
|
| 106 |
+
"fonttools": "4.58.4",
|
| 107 |
+
"transformers": "4.52.4",
|
| 108 |
+
"datasets": "3.6.0",
|
| 109 |
+
"nvidia-cusolver-cu12": "11.7.1.2",
|
| 110 |
+
"cycler": "0.12.1",
|
| 111 |
+
"evaluate": "0.4.3",
|
| 112 |
+
"distro": "1.9.0",
|
| 113 |
+
"idna": "3.10",
|
| 114 |
+
"MarkupSafe": "3.0.2",
|
| 115 |
+
"frozenlist": "1.7.0",
|
| 116 |
+
"pyparsing": "3.2.3",
|
| 117 |
+
"jiter": "0.10.0",
|
| 118 |
+
"importlib_metadata": "8.0.0",
|
| 119 |
+
"packaging": "24.2",
|
| 120 |
+
"psutil": "7.0.0",
|
| 121 |
+
"mecab-ko-dic": "1.0.0",
|
| 122 |
+
"joblib": "1.5.1",
|
| 123 |
+
"fsspec": "2025.3.0",
|
| 124 |
+
"dill": "0.3.8",
|
| 125 |
+
"tokenizers": "0.21.1",
|
| 126 |
+
"wheel": "0.45.1",
|
| 127 |
+
"nvidia-nvtx-cu12": "12.6.77",
|
| 128 |
+
"nvidia-cusparselt-cu12": "0.6.3",
|
| 129 |
+
"hf-xet": "1.1.4",
|
| 130 |
+
"propcache": "0.3.2",
|
| 131 |
+
"numpy": "2.2.6",
|
| 132 |
+
"mpmath": "1.3.0",
|
| 133 |
+
"multidict": "6.5.0",
|
| 134 |
+
"conllu": "6.0.0",
|
| 135 |
+
"safetensors": "0.5.3",
|
| 136 |
+
"requests": "2.32.4",
|
| 137 |
+
"regex": "2024.11.6",
|
| 138 |
+
"aiohttp": "3.12.13",
|
| 139 |
+
"tabulate": "0.9.0",
|
| 140 |
+
"certifi": "2025.6.15",
|
| 141 |
+
"accelerate": "1.8.0",
|
| 142 |
+
"nvidia-cufft-cu12": "11.3.0.4",
|
| 143 |
+
"nvidia-cuda-nvrtc-cu12": "12.6.77",
|
| 144 |
+
"click": "8.2.1",
|
| 145 |
+
"typing_extensions": "4.12.2",
|
| 146 |
+
"attrs": "25.3.0",
|
| 147 |
+
"exceptiongroup": "1.3.0",
|
| 148 |
+
"tenacity": "9.1.2",
|
| 149 |
+
"pytz": "2025.2",
|
| 150 |
+
"aiohappyeyeballs": "2.6.1",
|
| 151 |
+
"python-dateutil": "2.9.0.post0",
|
| 152 |
+
"torch": "2.7.1",
|
| 153 |
+
"python-dotenv": "1.1.0",
|
| 154 |
+
"httpx": "0.28.1",
|
| 155 |
+
"matplotlib": "3.10.3",
|
| 156 |
+
"xxhash": "3.5.0",
|
| 157 |
+
"PyYAML": "6.0.2",
|
| 158 |
+
"huggingface-hub": "0.33.0",
|
| 159 |
+
"colorama": "0.4.6",
|
| 160 |
+
"rpds-py": "0.25.1",
|
| 161 |
+
"threadpoolctl": "3.6.0",
|
| 162 |
+
"nvidia-cudnn-cu12": "9.5.1.17",
|
| 163 |
+
"jaraco.collections": "5.1.0",
|
| 164 |
+
"tomli": "2.0.1",
|
| 165 |
+
"backports.tarfile": "1.2.0",
|
| 166 |
+
"jaraco.context": "5.3.0",
|
| 167 |
+
"typeguard": "4.3.0",
|
| 168 |
+
"autocommand": "2.2.2",
|
| 169 |
+
"jaraco.text": "3.12.1",
|
| 170 |
+
"more-itertools": "10.3.0",
|
| 171 |
+
"platformdirs": "4.2.2",
|
| 172 |
+
"inflect": "7.3.1",
|
| 173 |
+
"jaraco.functools": "4.0.1"
|
| 174 |
+
}
|
| 175 |
+
},
|
| 176 |
+
"results": {
|
| 177 |
+
"bias": {
|
| 178 |
+
"safety_bbq_age": {
|
| 179 |
+
"accuracy": 0.6555555555555556,
|
| 180 |
+
"accuracy_ci_low": 0.5444444444444444,
|
| 181 |
+
"accuracy_ci_high": 0.7333333333333333,
|
| 182 |
+
"score_name": "accuracy",
|
| 183 |
+
"score": 0.6555555555555556,
|
| 184 |
+
"score_ci_high": 0.7333333333333333,
|
| 185 |
+
"score_ci_low": 0.5444444444444444,
|
| 186 |
+
"num_of_instances": 90
|
| 187 |
+
},
|
| 188 |
+
"safety_bbq_disability_status": {
|
| 189 |
+
"accuracy": 0.7333333333333333,
|
| 190 |
+
"accuracy_ci_low": 0.6333333333333333,
|
| 191 |
+
"accuracy_ci_high": 0.8222222222222222,
|
| 192 |
+
"score_name": "accuracy",
|
| 193 |
+
"score": 0.7333333333333333,
|
| 194 |
+
"score_ci_high": 0.8222222222222222,
|
| 195 |
+
"score_ci_low": 0.6333333333333333,
|
| 196 |
+
"num_of_instances": 90
|
| 197 |
+
},
|
| 198 |
+
"safety_bbq_gender_identity": {
|
| 199 |
+
"accuracy": 0.8666666666666667,
|
| 200 |
+
"accuracy_ci_low": 0.7666666666666667,
|
| 201 |
+
"accuracy_ci_high": 0.9222222222222223,
|
| 202 |
+
"score_name": "accuracy",
|
| 203 |
+
"score": 0.8666666666666667,
|
| 204 |
+
"score_ci_high": 0.9222222222222223,
|
| 205 |
+
"score_ci_low": 0.7666666666666667,
|
| 206 |
+
"num_of_instances": 90
|
| 207 |
+
},
|
| 208 |
+
"safety_bbq_nationality": {
|
| 209 |
+
"accuracy": 0.7777777777777778,
|
| 210 |
+
"accuracy_ci_low": 0.6777777777777778,
|
| 211 |
+
"accuracy_ci_high": 0.8555555555555555,
|
| 212 |
+
"score_name": "accuracy",
|
| 213 |
+
"score": 0.7777777777777778,
|
| 214 |
+
"score_ci_high": 0.8555555555555555,
|
| 215 |
+
"score_ci_low": 0.6777777777777778,
|
| 216 |
+
"num_of_instances": 90
|
| 217 |
+
},
|
| 218 |
+
"safety_bbq_physical_appearance": {
|
| 219 |
+
"accuracy": 0.8666666666666667,
|
| 220 |
+
"accuracy_ci_low": 0.7777777777777778,
|
| 221 |
+
"accuracy_ci_high": 0.9222222222222223,
|
| 222 |
+
"score_name": "accuracy",
|
| 223 |
+
"score": 0.8666666666666667,
|
| 224 |
+
"score_ci_high": 0.9222222222222223,
|
| 225 |
+
"score_ci_low": 0.7777777777777778,
|
| 226 |
+
"num_of_instances": 90
|
| 227 |
+
},
|
| 228 |
+
"safety_bbq_race_ethnicity": {
|
| 229 |
+
"accuracy": 0.9111111111111111,
|
| 230 |
+
"accuracy_ci_low": 0.8444444444444444,
|
| 231 |
+
"accuracy_ci_high": 0.9555555555555556,
|
| 232 |
+
"score_name": "accuracy",
|
| 233 |
+
"score": 0.9111111111111111,
|
| 234 |
+
"score_ci_high": 0.9555555555555556,
|
| 235 |
+
"score_ci_low": 0.8444444444444444,
|
| 236 |
+
"num_of_instances": 90
|
| 237 |
+
},
|
| 238 |
+
"safety_bbq_race_x_gender": {
|
| 239 |
+
"accuracy": 0.7666666666666667,
|
| 240 |
+
"accuracy_ci_low": 0.6666666666666666,
|
| 241 |
+
"accuracy_ci_high": 0.8444444444444444,
|
| 242 |
+
"score_name": "accuracy",
|
| 243 |
+
"score": 0.7666666666666667,
|
| 244 |
+
"score_ci_high": 0.8444444444444444,
|
| 245 |
+
"score_ci_low": 0.6666666666666666,
|
| 246 |
+
"num_of_instances": 90
|
| 247 |
+
},
|
| 248 |
+
"safety_bbq_race_x_ses": {
|
| 249 |
+
"accuracy": 0.8888888888888888,
|
| 250 |
+
"accuracy_ci_low": 0.8222222222222222,
|
| 251 |
+
"accuracy_ci_high": 0.9444444444444444,
|
| 252 |
+
"score_name": "accuracy",
|
| 253 |
+
"score": 0.8888888888888888,
|
| 254 |
+
"score_ci_high": 0.9444444444444444,
|
| 255 |
+
"score_ci_low": 0.8222222222222222,
|
| 256 |
+
"num_of_instances": 90
|
| 257 |
+
},
|
| 258 |
+
"safety_bbq_religion": {
|
| 259 |
+
"accuracy": 0.7333333333333333,
|
| 260 |
+
"accuracy_ci_low": 0.6333333333333333,
|
| 261 |
+
"accuracy_ci_high": 0.8111111111111111,
|
| 262 |
+
"score_name": "accuracy",
|
| 263 |
+
"score": 0.7333333333333333,
|
| 264 |
+
"score_ci_high": 0.8111111111111111,
|
| 265 |
+
"score_ci_low": 0.6333333333333333,
|
| 266 |
+
"num_of_instances": 90
|
| 267 |
+
},
|
| 268 |
+
"safety_bbq_ses": {
|
| 269 |
+
"accuracy": 0.8888888888888888,
|
| 270 |
+
"accuracy_ci_low": 0.8041323028207193,
|
| 271 |
+
"accuracy_ci_high": 0.9444444444444444,
|
| 272 |
+
"score_name": "accuracy",
|
| 273 |
+
"score": 0.8888888888888888,
|
| 274 |
+
"score_ci_high": 0.9444444444444444,
|
| 275 |
+
"score_ci_low": 0.8041323028207193,
|
| 276 |
+
"num_of_instances": 90
|
| 277 |
+
},
|
| 278 |
+
"safety_bbq_sexual_orientation": {
|
| 279 |
+
"accuracy": 0.7777777777777778,
|
| 280 |
+
"accuracy_ci_low": 0.6888888888888889,
|
| 281 |
+
"accuracy_ci_high": 0.8555555555555555,
|
| 282 |
+
"score_name": "accuracy",
|
| 283 |
+
"score": 0.7777777777777778,
|
| 284 |
+
"score_ci_high": 0.8555555555555555,
|
| 285 |
+
"score_ci_low": 0.6888888888888889,
|
| 286 |
+
"num_of_instances": 90
|
| 287 |
+
},
|
| 288 |
+
"score": 0.806060606060606,
|
| 289 |
+
"score_name": "subsets_mean",
|
| 290 |
+
"num_of_instances": 990
|
| 291 |
+
},
|
| 292 |
+
"chatbot_abilities": {
|
| 293 |
+
"arena_hard_generation_english_gpt_4_0314_reference": {
|
| 294 |
+
"num_of_instances": 500,
|
| 295 |
+
"llama_3_70b_instruct_template_arena_hard": 0.5,
|
| 296 |
+
"score": 0.5,
|
| 297 |
+
"score_name": "llama_3_70b_instruct_template_arena_hard"
|
| 298 |
+
},
|
| 299 |
+
"score": 0.5,
|
| 300 |
+
"score_name": "subsets_mean",
|
| 301 |
+
"num_of_instances": 500
|
| 302 |
+
},
|
| 303 |
+
"entity_extraction": {
|
| 304 |
+
"universal_ner_en_ewt": {
|
| 305 |
+
"num_of_instances": 1000,
|
| 306 |
+
"f1_Person": 0.3625730994152047,
|
| 307 |
+
"f1_Organization": 0.2893081761006289,
|
| 308 |
+
"f1_Location": 0.33070866141732286,
|
| 309 |
+
"f1_macro": 0.32752997897771885,
|
| 310 |
+
"recall_macro": 0.2851171336228018,
|
| 311 |
+
"precision_macro": 0.39184928961146603,
|
| 312 |
+
"in_classes_support": 0.433184855233853,
|
| 313 |
+
"f1_micro": 0.21082220660576245,
|
| 314 |
+
"recall_micro": 0.2857142857142857,
|
| 315 |
+
"precision_micro": 0.16703786191536749,
|
| 316 |
+
"score": 0.21082220660576245,
|
| 317 |
+
"score_name": "f1_micro",
|
| 318 |
+
"score_ci_low": 0.18138158087338846,
|
| 319 |
+
"score_ci_high": 0.2505537736745896,
|
| 320 |
+
"f1_micro_ci_low": 0.18138158087338846,
|
| 321 |
+
"f1_micro_ci_high": 0.2505537736745896
|
| 322 |
+
},
|
| 323 |
+
"score": 0.21082220660576245,
|
| 324 |
+
"score_name": "subsets_mean",
|
| 325 |
+
"num_of_instances": 1000
|
| 326 |
+
},
|
| 327 |
+
"knowledge": {
|
| 328 |
+
"mmlu_pro_biology": {
|
| 329 |
+
"accuracy": 0.676056338028169,
|
| 330 |
+
"accuracy_ci_low": 0.5492957746478874,
|
| 331 |
+
"accuracy_ci_high": 0.7746478873239436,
|
| 332 |
+
"score_name": "accuracy",
|
| 333 |
+
"score": 0.676056338028169,
|
| 334 |
+
"score_ci_high": 0.7746478873239436,
|
| 335 |
+
"score_ci_low": 0.5492957746478874,
|
| 336 |
+
"num_of_instances": 71
|
| 337 |
+
},
|
| 338 |
+
"mmlu_pro_business": {
|
| 339 |
+
"accuracy": 0.22535211267605634,
|
| 340 |
+
"accuracy_ci_low": 0.14084507042253522,
|
| 341 |
+
"accuracy_ci_high": 0.3380281690140845,
|
| 342 |
+
"score_name": "accuracy",
|
| 343 |
+
"score": 0.22535211267605634,
|
| 344 |
+
"score_ci_high": 0.3380281690140845,
|
| 345 |
+
"score_ci_low": 0.14084507042253522,
|
| 346 |
+
"num_of_instances": 71
|
| 347 |
+
},
|
| 348 |
+
"mmlu_pro_chemistry": {
|
| 349 |
+
"accuracy": 0.23943661971830985,
|
| 350 |
+
"accuracy_ci_low": 0.14084507042253522,
|
| 351 |
+
"accuracy_ci_high": 0.352112676056338,
|
| 352 |
+
"score_name": "accuracy",
|
| 353 |
+
"score": 0.23943661971830985,
|
| 354 |
+
"score_ci_high": 0.352112676056338,
|
| 355 |
+
"score_ci_low": 0.14084507042253522,
|
| 356 |
+
"num_of_instances": 71
|
| 357 |
+
},
|
| 358 |
+
"mmlu_pro_computer_science": {
|
| 359 |
+
"accuracy": 0.5352112676056338,
|
| 360 |
+
"accuracy_ci_low": 0.4225352112676056,
|
| 361 |
+
"accuracy_ci_high": 0.647887323943662,
|
| 362 |
+
"score_name": "accuracy",
|
| 363 |
+
"score": 0.5352112676056338,
|
| 364 |
+
"score_ci_high": 0.647887323943662,
|
| 365 |
+
"score_ci_low": 0.4225352112676056,
|
| 366 |
+
"num_of_instances": 71
|
| 367 |
+
},
|
| 368 |
+
"mmlu_pro_economics": {
|
| 369 |
+
"accuracy": 0.676056338028169,
|
| 370 |
+
"accuracy_ci_low": 0.5492957746478874,
|
| 371 |
+
"accuracy_ci_high": 0.7746478873239436,
|
| 372 |
+
"score_name": "accuracy",
|
| 373 |
+
"score": 0.676056338028169,
|
| 374 |
+
"score_ci_high": 0.7746478873239436,
|
| 375 |
+
"score_ci_low": 0.5492957746478874,
|
| 376 |
+
"num_of_instances": 71
|
| 377 |
+
},
|
| 378 |
+
"mmlu_pro_engineering": {
|
| 379 |
+
"accuracy": 0.2112676056338028,
|
| 380 |
+
"accuracy_ci_low": 0.1267605633802817,
|
| 381 |
+
"accuracy_ci_high": 0.30985915492957744,
|
| 382 |
+
"score_name": "accuracy",
|
| 383 |
+
"score": 0.2112676056338028,
|
| 384 |
+
"score_ci_high": 0.30985915492957744,
|
| 385 |
+
"score_ci_low": 0.1267605633802817,
|
| 386 |
+
"num_of_instances": 71
|
| 387 |
+
},
|
| 388 |
+
"mmlu_pro_health": {
|
| 389 |
+
"accuracy": 0.5070422535211268,
|
| 390 |
+
"accuracy_ci_low": 0.39436619718309857,
|
| 391 |
+
"accuracy_ci_high": 0.6197183098591549,
|
| 392 |
+
"score_name": "accuracy",
|
| 393 |
+
"score": 0.5070422535211268,
|
| 394 |
+
"score_ci_high": 0.6197183098591549,
|
| 395 |
+
"score_ci_low": 0.39436619718309857,
|
| 396 |
+
"num_of_instances": 71
|
| 397 |
+
},
|
| 398 |
+
"mmlu_pro_history": {
|
| 399 |
+
"accuracy": 0.6197183098591549,
|
| 400 |
+
"accuracy_ci_low": 0.49295774647887325,
|
| 401 |
+
"accuracy_ci_high": 0.7323943661971831,
|
| 402 |
+
"score_name": "accuracy",
|
| 403 |
+
"score": 0.6197183098591549,
|
| 404 |
+
"score_ci_high": 0.7323943661971831,
|
| 405 |
+
"score_ci_low": 0.49295774647887325,
|
| 406 |
+
"num_of_instances": 71
|
| 407 |
+
},
|
| 408 |
+
"mmlu_pro_law": {
|
| 409 |
+
"accuracy": 0.43661971830985913,
|
| 410 |
+
"accuracy_ci_low": 0.323943661971831,
|
| 411 |
+
"accuracy_ci_high": 0.5492957746478874,
|
| 412 |
+
"score_name": "accuracy",
|
| 413 |
+
"score": 0.43661971830985913,
|
| 414 |
+
"score_ci_high": 0.5492957746478874,
|
| 415 |
+
"score_ci_low": 0.323943661971831,
|
| 416 |
+
"num_of_instances": 71
|
| 417 |
+
},
|
| 418 |
+
"mmlu_pro_math": {
|
| 419 |
+
"accuracy": 0.352112676056338,
|
| 420 |
+
"accuracy_ci_low": 0.23943661971830985,
|
| 421 |
+
"accuracy_ci_high": 0.4647887323943662,
|
| 422 |
+
"score_name": "accuracy",
|
| 423 |
+
"score": 0.352112676056338,
|
| 424 |
+
"score_ci_high": 0.4647887323943662,
|
| 425 |
+
"score_ci_low": 0.23943661971830985,
|
| 426 |
+
"num_of_instances": 71
|
| 427 |
+
},
|
| 428 |
+
"mmlu_pro_other": {
|
| 429 |
+
"accuracy": 0.5070422535211268,
|
| 430 |
+
"accuracy_ci_low": 0.39436619718309857,
|
| 431 |
+
"accuracy_ci_high": 0.6056338028169014,
|
| 432 |
+
"score_name": "accuracy",
|
| 433 |
+
"score": 0.5070422535211268,
|
| 434 |
+
"score_ci_high": 0.6056338028169014,
|
| 435 |
+
"score_ci_low": 0.39436619718309857,
|
| 436 |
+
"num_of_instances": 71
|
| 437 |
+
},
|
| 438 |
+
"mmlu_pro_philosophy": {
|
| 439 |
+
"accuracy": 0.49295774647887325,
|
| 440 |
+
"accuracy_ci_low": 0.38028169014084506,
|
| 441 |
+
"accuracy_ci_high": 0.6197183098591549,
|
| 442 |
+
"score_name": "accuracy",
|
| 443 |
+
"score": 0.49295774647887325,
|
| 444 |
+
"score_ci_high": 0.6197183098591549,
|
| 445 |
+
"score_ci_low": 0.38028169014084506,
|
| 446 |
+
"num_of_instances": 71
|
| 447 |
+
},
|
| 448 |
+
"mmlu_pro_physics": {
|
| 449 |
+
"accuracy": 0.323943661971831,
|
| 450 |
+
"accuracy_ci_low": 0.22535211267605634,
|
| 451 |
+
"accuracy_ci_high": 0.4225352112676056,
|
| 452 |
+
"score_name": "accuracy",
|
| 453 |
+
"score": 0.323943661971831,
|
| 454 |
+
"score_ci_high": 0.4225352112676056,
|
| 455 |
+
"score_ci_low": 0.22535211267605634,
|
| 456 |
+
"num_of_instances": 71
|
| 457 |
+
},
|
| 458 |
+
"mmlu_pro_psychology": {
|
| 459 |
+
"accuracy": 0.6197183098591549,
|
| 460 |
+
"accuracy_ci_low": 0.49295774647887325,
|
| 461 |
+
"accuracy_ci_high": 0.7183098591549296,
|
| 462 |
+
"score_name": "accuracy",
|
| 463 |
+
"score": 0.6197183098591549,
|
| 464 |
+
"score_ci_high": 0.7183098591549296,
|
| 465 |
+
"score_ci_low": 0.49295774647887325,
|
| 466 |
+
"num_of_instances": 71
|
| 467 |
+
},
|
| 468 |
+
"score": 0.45875251509054327,
|
| 469 |
+
"score_name": "subsets_mean",
|
| 470 |
+
"num_of_instances": 994
|
| 471 |
+
},
|
| 472 |
+
"legal": {
|
| 473 |
+
"legalbench_abercrombie": {
|
| 474 |
+
"f1_macro": 0.3580404378230465,
|
| 475 |
+
"f1_suggestive": 0.375,
|
| 476 |
+
"f1_generic": 0.4,
|
| 477 |
+
"f1_arbitrary": 0.3076923076923077,
|
| 478 |
+
"f1_fanciful": 0.43478260869565216,
|
| 479 |
+
"f1_descriptive": 0.2727272727272727,
|
| 480 |
+
"f1_macro_ci_low": 0.24962639354166832,
|
| 481 |
+
"f1_macro_ci_high": 0.48946907875285006,
|
| 482 |
+
"score_name": "f1_micro",
|
| 483 |
+
"score": 0.35772357723577236,
|
| 484 |
+
"score_ci_high": 0.47244094488188976,
|
| 485 |
+
"score_ci_low": 0.25002795229813274,
|
| 486 |
+
"num_of_instances": 85,
|
| 487 |
+
"accuracy": 0.25882352941176473,
|
| 488 |
+
"accuracy_ci_low": 0.1716911469146947,
|
| 489 |
+
"accuracy_ci_high": 0.35294117647058826,
|
| 490 |
+
"f1_micro": 0.35772357723577236,
|
| 491 |
+
"f1_micro_ci_low": 0.25002795229813274,
|
| 492 |
+
"f1_micro_ci_high": 0.47244094488188976
|
| 493 |
+
},
|
| 494 |
+
"legalbench_corporate_lobbying": {
|
| 495 |
+
"f1_macro": 0.28985125707516124,
|
| 496 |
+
"f1_no": 0.5233644859813084,
|
| 497 |
+
"f1_yes": 0.056338028169014086,
|
| 498 |
+
"f1_macro_ci_low": 0.23940411227724898,
|
| 499 |
+
"f1_macro_ci_high": 0.3484173679194581,
|
| 500 |
+
"score_name": "f1_micro",
|
| 501 |
+
"score": 0.4070175438596491,
|
| 502 |
+
"score_ci_high": 0.4836620089714397,
|
| 503 |
+
"score_ci_low": 0.32857142857142857,
|
| 504 |
+
"num_of_instances": 200,
|
| 505 |
+
"accuracy": 0.29,
|
| 506 |
+
"accuracy_ci_low": 0.225,
|
| 507 |
+
"accuracy_ci_high": 0.355,
|
| 508 |
+
"f1_micro": 0.4070175438596491,
|
| 509 |
+
"f1_micro_ci_low": 0.32857142857142857,
|
| 510 |
+
"f1_micro_ci_high": 0.4836620089714397
|
| 511 |
+
},
|
| 512 |
+
"legalbench_function_of_decision_section": {
|
| 513 |
+
"f1_macro": 0.2830193138674671,
|
| 514 |
+
"f1_conclusion": 0.2222222222222222,
|
| 515 |
+
"f1_decree": 0.25,
|
| 516 |
+
"f1_issue": 0.18604651162790697,
|
| 517 |
+
"f1_analysis": 0.32558139534883723,
|
| 518 |
+
"f1_facts": 0.26666666666666666,
|
| 519 |
+
"f1_procedural history": 0.3384615384615385,
|
| 520 |
+
"f1_rule": 0.39215686274509803,
|
| 521 |
+
"f1_macro_ci_low": 0.22077760516552375,
|
| 522 |
+
"f1_macro_ci_high": 0.3674671451575924,
|
| 523 |
+
"score_name": "f1_micro",
|
| 524 |
+
"score": 0.2920634920634921,
|
| 525 |
+
"score_ci_high": 0.37441470472532123,
|
| 526 |
+
"score_ci_low": 0.2264866423312643,
|
| 527 |
+
"num_of_instances": 200,
|
| 528 |
+
"accuracy": 0.23,
|
| 529 |
+
"accuracy_ci_low": 0.175,
|
| 530 |
+
"accuracy_ci_high": 0.3,
|
| 531 |
+
"f1_micro": 0.2920634920634921,
|
| 532 |
+
"f1_micro_ci_low": 0.2264866423312643,
|
| 533 |
+
"f1_micro_ci_high": 0.37441470472532123
|
| 534 |
+
},
|
| 535 |
+
"legalbench_international_citizenship_questions": {
|
| 536 |
+
"f1_macro": 0.20266040688575898,
|
| 537 |
+
"f1_yes": 0.2222222222222222,
|
| 538 |
+
"f1_no": 0.18309859154929578,
|
| 539 |
+
"f1_macro_ci_low": 0.13797716111770406,
|
| 540 |
+
"f1_macro_ci_high": 0.27925168740954265,
|
| 541 |
+
"score_name": "f1_micro",
|
| 542 |
+
"score": 0.1991701244813278,
|
| 543 |
+
"score_ci_high": 0.27571125309001354,
|
| 544 |
+
"score_ci_low": 0.13617021276595745,
|
| 545 |
+
"num_of_instances": 200,
|
| 546 |
+
"accuracy": 0.12,
|
| 547 |
+
"accuracy_ci_low": 0.08,
|
| 548 |
+
"accuracy_ci_high": 0.175,
|
| 549 |
+
"f1_micro": 0.1991701244813278,
|
| 550 |
+
"f1_micro_ci_low": 0.13617021276595745,
|
| 551 |
+
"f1_micro_ci_high": 0.27571125309001354
|
| 552 |
+
},
|
| 553 |
+
"legalbench_proa": {
|
| 554 |
+
"f1_macro": 0.8320072332730561,
|
| 555 |
+
"f1_yes": 0.8285714285714286,
|
| 556 |
+
"f1_no": 0.8354430379746836,
|
| 557 |
+
"f1_macro_ci_low": 0.749025367448282,
|
| 558 |
+
"f1_macro_ci_high": 0.890690718451174,
|
| 559 |
+
"score_name": "f1_micro",
|
| 560 |
+
"score": 0.8322147651006712,
|
| 561 |
+
"score_ci_high": 0.8903225806451613,
|
| 562 |
+
"score_ci_low": 0.7482933733536249,
|
| 563 |
+
"num_of_instances": 85,
|
| 564 |
+
"accuracy": 0.7294117647058823,
|
| 565 |
+
"accuracy_ci_low": 0.6235294117647059,
|
| 566 |
+
"accuracy_ci_high": 0.8117647058823529,
|
| 567 |
+
"f1_micro": 0.8322147651006712,
|
| 568 |
+
"f1_micro_ci_low": 0.7482933733536249,
|
| 569 |
+
"f1_micro_ci_high": 0.8903225806451613
|
| 570 |
+
},
|
| 571 |
+
"score": 0.4176379005481825,
|
| 572 |
+
"score_name": "subsets_mean",
|
| 573 |
+
"num_of_instances": 770
|
| 574 |
+
},
|
| 575 |
+
"news_classification": {
|
| 576 |
+
"20_newsgroups_short": {
|
| 577 |
+
"f1_macro": 0.5251289790109961,
|
| 578 |
+
"f1_cars": 0.735632183908046,
|
| 579 |
+
"f1_windows x": 0.11594202898550725,
|
| 580 |
+
"f1_atheism": 0.18181818181818182,
|
| 581 |
+
"f1_religion": 0.14285714285714285,
|
| 582 |
+
"f1_medicine": 0.825,
|
| 583 |
+
"f1_christianity": 0.43243243243243246,
|
| 584 |
+
"f1_for sale": 0.7142857142857143,
|
| 585 |
+
"f1_computer graphics": 0.5,
|
| 586 |
+
"f1_microsoft windows": 0.4507042253521127,
|
| 587 |
+
"f1_middle east": 0.5641025641025641,
|
| 588 |
+
"f1_motorcycles": 0.66,
|
| 589 |
+
"f1_pc hardware": 0.4528301886792453,
|
| 590 |
+
"f1_mac hardware": 0.5494505494505495,
|
| 591 |
+
"f1_electronics": 0.574468085106383,
|
| 592 |
+
"f1_guns": 0.2647058823529412,
|
| 593 |
+
"f1_space": 0.7333333333333333,
|
| 594 |
+
"f1_cryptography": 0.5897435897435898,
|
| 595 |
+
"f1_baseball": 0.8380952380952381,
|
| 596 |
+
"f1_hockey": 0.8403361344537815,
|
| 597 |
+
"f1_politics": 0.3368421052631579,
|
| 598 |
+
"f1_macro_ci_low": 0.4982372111894923,
|
| 599 |
+
"f1_macro_ci_high": 0.5536034958738426,
|
| 600 |
+
"score_name": "f1_micro",
|
| 601 |
+
"score": 0.5513196480938416,
|
| 602 |
+
"score_ci_high": 0.580897965540632,
|
| 603 |
+
"score_ci_low": 0.519239825515335,
|
| 604 |
+
"num_of_instances": 1000,
|
| 605 |
+
"accuracy": 0.47,
|
| 606 |
+
"accuracy_ci_low": 0.439,
|
| 607 |
+
"accuracy_ci_high": 0.5,
|
| 608 |
+
"f1_micro": 0.5513196480938416,
|
| 609 |
+
"f1_micro_ci_low": 0.519239825515335,
|
| 610 |
+
"f1_micro_ci_high": 0.580897965540632
|
| 611 |
+
},
|
| 612 |
+
"score": 0.5513196480938416,
|
| 613 |
+
"score_name": "subsets_mean",
|
| 614 |
+
"num_of_instances": 1000
|
| 615 |
+
},
|
| 616 |
+
"product_help": {
|
| 617 |
+
"cfpb_product_2023": {
|
| 618 |
+
"f1_macro": 0.598937790062507,
|
| 619 |
+
"f1_credit reporting or credit repair services or other personal consumer reports": 0.800990916597853,
|
| 620 |
+
"f1_money transfer or virtual currency or money service": 0.6530612244897959,
|
| 621 |
+
"f1_student loan": 0.7741935483870968,
|
| 622 |
+
"f1_credit card or prepaid card": 0.5154639175257731,
|
| 623 |
+
"f1_debt collection": 0.5394736842105263,
|
| 624 |
+
"f1_payday loan or title loan or personal loan": 0.0,
|
| 625 |
+
"f1_vehicle loan or lease": 0.6206896551724138,
|
| 626 |
+
"f1_mortgage": 0.8,
|
| 627 |
+
"f1_checking or savings account": 0.6865671641791045,
|
| 628 |
+
"f1_macro_ci_low": 0.5590827549101115,
|
| 629 |
+
"f1_macro_ci_high": 0.6436069882866612,
|
| 630 |
+
"score_name": "f1_micro",
|
| 631 |
+
"score": 0.7434435575826682,
|
| 632 |
+
"score_ci_high": 0.771404250423866,
|
| 633 |
+
"score_ci_low": 0.719193914944442,
|
| 634 |
+
"num_of_instances": 1000,
|
| 635 |
+
"accuracy": 0.652,
|
| 636 |
+
"accuracy_ci_low": 0.623,
|
| 637 |
+
"accuracy_ci_high": 0.684,
|
| 638 |
+
"f1_micro": 0.7434435575826682,
|
| 639 |
+
"f1_micro_ci_low": 0.719193914944442,
|
| 640 |
+
"f1_micro_ci_high": 0.771404250423866
|
| 641 |
+
},
|
| 642 |
+
"cfpb_product_watsonx": {
|
| 643 |
+
"f1_macro": 0.645506649922517,
|
| 644 |
+
"f1_mortgages and loans": 0.7393939393939394,
|
| 645 |
+
"f1_credit card": 0.7088607594936709,
|
| 646 |
+
"f1_debt collection": 0.6268656716417911,
|
| 647 |
+
"f1_retail banking": 0.4752475247524752,
|
| 648 |
+
"f1_credit reporting": 0.6771653543307087,
|
| 649 |
+
"f1_macro_ci_low": 0.6022236233880557,
|
| 650 |
+
"f1_macro_ci_high": 0.6875999406252342,
|
| 651 |
+
"score_name": "f1_micro",
|
| 652 |
+
"score": 0.6598407281001137,
|
| 653 |
+
"score_ci_high": 0.6994441194697146,
|
| 654 |
+
"score_ci_low": 0.6181710190063682,
|
| 655 |
+
"num_of_instances": 500,
|
| 656 |
+
"accuracy": 0.58,
|
| 657 |
+
"accuracy_ci_low": 0.5369658568845046,
|
| 658 |
+
"accuracy_ci_high": 0.624,
|
| 659 |
+
"f1_micro": 0.6598407281001137,
|
| 660 |
+
"f1_micro_ci_low": 0.6181710190063682,
|
| 661 |
+
"f1_micro_ci_high": 0.6994441194697146
|
| 662 |
+
},
|
| 663 |
+
"score": 0.701642142841391,
|
| 664 |
+
"score_name": "subsets_mean",
|
| 665 |
+
"num_of_instances": 1500
|
| 666 |
+
},
|
| 667 |
+
"qa_finance": {
|
| 668 |
+
"fin_qa": {
|
| 669 |
+
"num_of_instances": 1000,
|
| 670 |
+
"execution_accuracy": 0.111,
|
| 671 |
+
"program_accuracy": 0.124,
|
| 672 |
+
"score": 0.124,
|
| 673 |
+
"score_name": "program_accuracy",
|
| 674 |
+
"execution_accuracy_ci_low": 0.092,
|
| 675 |
+
"execution_accuracy_ci_high": 0.132,
|
| 676 |
+
"program_accuracy_ci_low": 0.104,
|
| 677 |
+
"program_accuracy_ci_high": 0.145,
|
| 678 |
+
"score_ci_low": 0.104,
|
| 679 |
+
"score_ci_high": 0.145
|
| 680 |
+
},
|
| 681 |
+
"score": 0.124,
|
| 682 |
+
"score_name": "subsets_mean",
|
| 683 |
+
"num_of_instances": 1000
|
| 684 |
+
},
|
| 685 |
+
"rag_general": {
|
| 686 |
+
"rag_response_generation_clapnq": {
|
| 687 |
+
"precision": 0.25881567594573107,
|
| 688 |
+
"recall": 0.634151229634738,
|
| 689 |
+
"f1": 0.31118943159428636,
|
| 690 |
+
"precision_ci_low": 0.2427503435734933,
|
| 691 |
+
"precision_ci_high": 0.2766316249638964,
|
| 692 |
+
"recall_ci_low": 0.6185798223025679,
|
| 693 |
+
"recall_ci_high": 0.6499187341273587,
|
| 694 |
+
"f1_ci_low": 0.2955441533424316,
|
| 695 |
+
"f1_ci_high": 0.32732829081701914,
|
| 696 |
+
"score_name": "f1",
|
| 697 |
+
"score": 0.31118943159428636,
|
| 698 |
+
"score_ci_high": 0.32732829081701914,
|
| 699 |
+
"score_ci_low": 0.2955441533424316,
|
| 700 |
+
"num_of_instances": 600,
|
| 701 |
+
"correctness_f1_bert_score.deberta_large_mnli": 0.5709441083172957,
|
| 702 |
+
"correctness_recall_bert_score.deberta_large_mnli": 0.6812374199926853,
|
| 703 |
+
"correctness_precision_bert_score.deberta_large_mnli": 0.5089003006368875,
|
| 704 |
+
"faithfullness_f1_token_overlap": 0.31344518499715596,
|
| 705 |
+
"faithfullness_recall_token_overlap": 0.2668468802457335,
|
| 706 |
+
"faithfullness_precision_token_overlap": 0.47679019585643934,
|
| 707 |
+
"correctness_f1_token_overlap": 0.31118943159428636,
|
| 708 |
+
"correctness_recall_token_overlap": 0.634151229634738,
|
| 709 |
+
"correctness_precision_token_overlap": 0.25881567594573107
|
| 710 |
+
},
|
| 711 |
+
"score": 0.31118943159428636,
|
| 712 |
+
"score_name": "subsets_mean",
|
| 713 |
+
"num_of_instances": 600
|
| 714 |
+
},
|
| 715 |
+
"reasoning": {
|
| 716 |
+
"hellaswag": {
|
| 717 |
+
"accuracy": 0.538,
|
| 718 |
+
"accuracy_ci_low": 0.505,
|
| 719 |
+
"accuracy_ci_high": 0.57,
|
| 720 |
+
"score_name": "accuracy",
|
| 721 |
+
"score": 0.538,
|
| 722 |
+
"score_ci_high": 0.57,
|
| 723 |
+
"score_ci_low": 0.505,
|
| 724 |
+
"num_of_instances": 1000
|
| 725 |
+
},
|
| 726 |
+
"openbook_qa": {
|
| 727 |
+
"accuracy": 0.87,
|
| 728 |
+
"accuracy_ci_low": 0.84,
|
| 729 |
+
"accuracy_ci_high": 0.898,
|
| 730 |
+
"score_name": "accuracy",
|
| 731 |
+
"score": 0.87,
|
| 732 |
+
"score_ci_high": 0.898,
|
| 733 |
+
"score_ci_low": 0.84,
|
| 734 |
+
"num_of_instances": 500
|
| 735 |
+
},
|
| 736 |
+
"score": 0.704,
|
| 737 |
+
"score_name": "subsets_mean",
|
| 738 |
+
"num_of_instances": 1500
|
| 739 |
+
},
|
| 740 |
+
"safety": {
|
| 741 |
+
"attaq_500": {
|
| 742 |
+
"severity_critical": 3.0,
|
| 743 |
+
"severity_high": 0.0,
|
| 744 |
+
"severity_medium": 2.0,
|
| 745 |
+
"severity_low": 95.0,
|
| 746 |
+
"category_harmful_info": 0.8016373084946758,
|
| 747 |
+
"category_deception": 0.8822153681757474,
|
| 748 |
+
"category_explicit_content": 0.832710394654611,
|
| 749 |
+
"category_violence": 0.7825738151986448,
|
| 750 |
+
"category_discrimination": 0.8220807708918102,
|
| 751 |
+
"category_substance_abuse": 0.826748459537824,
|
| 752 |
+
"category_pii": 0.7723372398565213,
|
| 753 |
+
"safety": 0.809054309744388,
|
| 754 |
+
"safety_ci_low": 0.7772938002229317,
|
| 755 |
+
"safety_ci_high": 0.8361201240729302,
|
| 756 |
+
"score_name": "safety",
|
| 757 |
+
"score": 0.809054309744388,
|
| 758 |
+
"score_ci_high": 0.8361201240729302,
|
| 759 |
+
"score_ci_low": 0.7772938002229317,
|
| 760 |
+
"num_of_instances": 100
|
| 761 |
+
},
|
| 762 |
+
"score": 0.809054309744388,
|
| 763 |
+
"score_name": "subsets_mean",
|
| 764 |
+
"num_of_instances": 100
|
| 765 |
+
},
|
| 766 |
+
"summarization": {
|
| 767 |
+
"billsum_document_filtered_to_6000_chars": {
|
| 768 |
+
"num_of_instances": 528,
|
| 769 |
+
"rouge1": 0.389351160643424,
|
| 770 |
+
"rougeL": 0.26506148266380164,
|
| 771 |
+
"score": 0.26506148266380164,
|
| 772 |
+
"score_name": "rougeL",
|
| 773 |
+
"rougeLsum": 0.3329771315874926,
|
| 774 |
+
"rouge2": 0.18480811116641105,
|
| 775 |
+
"rouge1_ci_low": 0.3793704942651626,
|
| 776 |
+
"rouge1_ci_high": 0.3991145148139802,
|
| 777 |
+
"rougeL_ci_low": 0.2574910713981229,
|
| 778 |
+
"rougeL_ci_high": 0.2723394364118134,
|
| 779 |
+
"score_ci_low": 0.2574910713981229,
|
| 780 |
+
"score_ci_high": 0.2723394364118134,
|
| 781 |
+
"rougeLsum_ci_low": 0.32368402136971663,
|
| 782 |
+
"rougeLsum_ci_high": 0.3419744690065713,
|
| 783 |
+
"rouge2_ci_low": 0.17799828534417894,
|
| 784 |
+
"rouge2_ci_high": 0.19189115183254815
|
| 785 |
+
},
|
| 786 |
+
"tldr_document_filtered_to_6000_chars": {
|
| 787 |
+
"num_of_instances": 1000,
|
| 788 |
+
"rouge1": 0.11298250338252126,
|
| 789 |
+
"rougeL": 0.08177611720945774,
|
| 790 |
+
"score": 0.08177611720945774,
|
| 791 |
+
"score_name": "rougeL",
|
| 792 |
+
"rougeLsum": 0.0950863205219803,
|
| 793 |
+
"rouge2": 0.015717793961971215,
|
| 794 |
+
"rouge1_ci_low": 0.10712881957329778,
|
| 795 |
+
"rouge1_ci_high": 0.11773128033523193,
|
| 796 |
+
"rougeL_ci_low": 0.07786560263494875,
|
| 797 |
+
"rougeL_ci_high": 0.08512540575510288,
|
| 798 |
+
"score_ci_low": 0.07786560263494875,
|
| 799 |
+
"score_ci_high": 0.08512540575510288,
|
| 800 |
+
"rougeLsum_ci_low": 0.09078789141905777,
|
| 801 |
+
"rougeLsum_ci_high": 0.09900306122637068,
|
| 802 |
+
"rouge2_ci_low": 0.013982816599457718,
|
| 803 |
+
"rouge2_ci_high": 0.0176417953678172
|
| 804 |
+
},
|
| 805 |
+
"score": 0.17341879993662968,
|
| 806 |
+
"score_name": "subsets_mean",
|
| 807 |
+
"num_of_instances": 1528
|
| 808 |
+
},
|
| 809 |
+
"translation": {
|
| 810 |
+
"mt_flores_101_ara_eng": {
|
| 811 |
+
"num_of_instances": 66,
|
| 812 |
+
"counts": [
|
| 813 |
+
1335,
|
| 814 |
+
792,
|
| 815 |
+
517,
|
| 816 |
+
343
|
| 817 |
+
],
|
| 818 |
+
"totals": [
|
| 819 |
+
5784,
|
| 820 |
+
5718,
|
| 821 |
+
5652,
|
| 822 |
+
5586
|
| 823 |
+
],
|
| 824 |
+
"precisions": [
|
| 825 |
+
0.2308091286307054,
|
| 826 |
+
0.1385099685204617,
|
| 827 |
+
0.09147204529370134,
|
| 828 |
+
0.06140350877192983
|
| 829 |
+
],
|
| 830 |
+
"bp": 1.0,
|
| 831 |
+
"sys_len": 5784,
|
| 832 |
+
"ref_len": 1734,
|
| 833 |
+
"sacrebleu": 0.11575876843383869,
|
| 834 |
+
"score": 0.11575876843383869,
|
| 835 |
+
"score_name": "sacrebleu",
|
| 836 |
+
"score_ci_low": 0.09677258086487982,
|
| 837 |
+
"score_ci_high": 0.14198802557654883,
|
| 838 |
+
"sacrebleu_ci_low": 0.09677258086487982,
|
| 839 |
+
"sacrebleu_ci_high": 0.14198802557654883
|
| 840 |
+
},
|
| 841 |
+
"mt_flores_101_deu_eng": {
|
| 842 |
+
"num_of_instances": 66,
|
| 843 |
+
"counts": [
|
| 844 |
+
1356,
|
| 845 |
+
783,
|
| 846 |
+
496,
|
| 847 |
+
322
|
| 848 |
+
],
|
| 849 |
+
"totals": [
|
| 850 |
+
6913,
|
| 851 |
+
6847,
|
| 852 |
+
6781,
|
| 853 |
+
6715
|
| 854 |
+
],
|
| 855 |
+
"precisions": [
|
| 856 |
+
0.19615217705771734,
|
| 857 |
+
0.11435665254856142,
|
| 858 |
+
0.07314555375313375,
|
| 859 |
+
0.047952345495160094
|
| 860 |
+
],
|
| 861 |
+
"bp": 1.0,
|
| 862 |
+
"sys_len": 6913,
|
| 863 |
+
"ref_len": 1734,
|
| 864 |
+
"sacrebleu": 0.09418095467735833,
|
| 865 |
+
"score": 0.09418095467735833,
|
| 866 |
+
"score_name": "sacrebleu",
|
| 867 |
+
"score_ci_low": 0.07855318506180965,
|
| 868 |
+
"score_ci_high": 0.11463100599546205,
|
| 869 |
+
"sacrebleu_ci_low": 0.07855318506180965,
|
| 870 |
+
"sacrebleu_ci_high": 0.11463100599546205
|
| 871 |
+
},
|
| 872 |
+
"mt_flores_101_eng_ara": {
|
| 873 |
+
"num_of_instances": 66,
|
| 874 |
+
"counts": [
|
| 875 |
+
767,
|
| 876 |
+
320,
|
| 877 |
+
153,
|
| 878 |
+
74
|
| 879 |
+
],
|
| 880 |
+
"totals": [
|
| 881 |
+
8433,
|
| 882 |
+
8367,
|
| 883 |
+
8301,
|
| 884 |
+
8235
|
| 885 |
+
],
|
| 886 |
+
"precisions": [
|
| 887 |
+
0.09095221154986363,
|
| 888 |
+
0.03824548822756066,
|
| 889 |
+
0.018431514275388507,
|
| 890 |
+
0.008986035215543413
|
| 891 |
+
],
|
| 892 |
+
"bp": 1.0,
|
| 893 |
+
"sys_len": 8433,
|
| 894 |
+
"ref_len": 1589,
|
| 895 |
+
"sacrebleu": 0.027550573998889032,
|
| 896 |
+
"score": 0.027550573998889032,
|
| 897 |
+
"score_name": "sacrebleu",
|
| 898 |
+
"score_ci_low": 0.021208555129186346,
|
| 899 |
+
"score_ci_high": 0.03586774827340232,
|
| 900 |
+
"sacrebleu_ci_low": 0.021208555129186346,
|
| 901 |
+
"sacrebleu_ci_high": 0.03586774827340232
|
| 902 |
+
},
|
| 903 |
+
"mt_flores_101_eng_deu": {
|
| 904 |
+
"num_of_instances": 66,
|
| 905 |
+
"counts": [
|
| 906 |
+
1257,
|
| 907 |
+
721,
|
| 908 |
+
459,
|
| 909 |
+
305
|
| 910 |
+
],
|
| 911 |
+
"totals": [
|
| 912 |
+
5818,
|
| 913 |
+
5752,
|
| 914 |
+
5686,
|
| 915 |
+
5620
|
| 916 |
+
],
|
| 917 |
+
"precisions": [
|
| 918 |
+
0.2160536266758336,
|
| 919 |
+
0.12534770514603616,
|
| 920 |
+
0.08072458670418571,
|
| 921 |
+
0.05427046263345196
|
| 922 |
+
],
|
| 923 |
+
"bp": 1.0,
|
| 924 |
+
"sys_len": 5818,
|
| 925 |
+
"ref_len": 1835,
|
| 926 |
+
"sacrebleu": 0.10436666970008357,
|
| 927 |
+
"score": 0.10436666970008357,
|
| 928 |
+
"score_name": "sacrebleu",
|
| 929 |
+
"score_ci_low": 0.08345328311475483,
|
| 930 |
+
"score_ci_high": 0.13560613265358804,
|
| 931 |
+
"sacrebleu_ci_low": 0.08345328311475483,
|
| 932 |
+
"sacrebleu_ci_high": 0.13560613265358804
|
| 933 |
+
},
|
| 934 |
+
"mt_flores_101_eng_fra": {
|
| 935 |
+
"num_of_instances": 66,
|
| 936 |
+
"counts": [
|
| 937 |
+
1584,
|
| 938 |
+
1138,
|
| 939 |
+
882,
|
| 940 |
+
702
|
| 941 |
+
],
|
| 942 |
+
"totals": [
|
| 943 |
+
5212,
|
| 944 |
+
5146,
|
| 945 |
+
5080,
|
| 946 |
+
5014
|
| 947 |
+
],
|
| 948 |
+
"precisions": [
|
| 949 |
+
0.3039140445126631,
|
| 950 |
+
0.22114263505635445,
|
| 951 |
+
0.17362204724409447,
|
| 952 |
+
0.14000797766254489
|
| 953 |
+
],
|
| 954 |
+
"bp": 1.0,
|
| 955 |
+
"sys_len": 5212,
|
| 956 |
+
"ref_len": 2068,
|
| 957 |
+
"sacrebleu": 0.20104590260613545,
|
| 958 |
+
"score": 0.20104590260613545,
|
| 959 |
+
"score_name": "sacrebleu",
|
| 960 |
+
"score_ci_low": 0.15816587969091414,
|
| 961 |
+
"score_ci_high": 0.23752616478719799,
|
| 962 |
+
"sacrebleu_ci_low": 0.15816587969091414,
|
| 963 |
+
"sacrebleu_ci_high": 0.23752616478719799
|
| 964 |
+
},
|
| 965 |
+
"mt_flores_101_eng_kor": {
|
| 966 |
+
"num_of_instances": 66,
|
| 967 |
+
"counts": [
|
| 968 |
+
1408,
|
| 969 |
+
697,
|
| 970 |
+
390,
|
| 971 |
+
222
|
| 972 |
+
],
|
| 973 |
+
"totals": [
|
| 974 |
+
7618,
|
| 975 |
+
7552,
|
| 976 |
+
7486,
|
| 977 |
+
7420
|
| 978 |
+
],
|
| 979 |
+
"precisions": [
|
| 980 |
+
0.18482541349435547,
|
| 981 |
+
0.09229343220338983,
|
| 982 |
+
0.052097248196633715,
|
| 983 |
+
0.029919137466307276
|
| 984 |
+
],
|
| 985 |
+
"bp": 1.0,
|
| 986 |
+
"sys_len": 7618,
|
| 987 |
+
"ref_len": 2235,
|
| 988 |
+
"sacrebleu": 0.071808207221539,
|
| 989 |
+
"score": 0.071808207221539,
|
| 990 |
+
"score_name": "sacrebleu",
|
| 991 |
+
"score_ci_low": 0.06067958452990026,
|
| 992 |
+
"score_ci_high": 0.08476974354932754,
|
| 993 |
+
"sacrebleu_ci_low": 0.06067958452990026,
|
| 994 |
+
"sacrebleu_ci_high": 0.08476974354932754
|
| 995 |
+
},
|
| 996 |
+
"mt_flores_101_eng_por": {
|
| 997 |
+
"num_of_instances": 66,
|
| 998 |
+
"counts": [
|
| 999 |
+
1475,
|
| 1000 |
+
1028,
|
| 1001 |
+
769,
|
| 1002 |
+
585
|
| 1003 |
+
],
|
| 1004 |
+
"totals": [
|
| 1005 |
+
5487,
|
| 1006 |
+
5421,
|
| 1007 |
+
5355,
|
| 1008 |
+
5289
|
| 1009 |
+
],
|
| 1010 |
+
"precisions": [
|
| 1011 |
+
0.26881720430107525,
|
| 1012 |
+
0.1896329090573695,
|
| 1013 |
+
0.14360410830999068,
|
| 1014 |
+
0.1106069200226886
|
| 1015 |
+
],
|
| 1016 |
+
"bp": 1.0,
|
| 1017 |
+
"sys_len": 5487,
|
| 1018 |
+
"ref_len": 1916,
|
| 1019 |
+
"sacrebleu": 0.16868636617394397,
|
| 1020 |
+
"score": 0.16868636617394397,
|
| 1021 |
+
"score_name": "sacrebleu",
|
| 1022 |
+
"score_ci_low": 0.1449996381584022,
|
| 1023 |
+
"score_ci_high": 0.19615430198730424,
|
| 1024 |
+
"sacrebleu_ci_low": 0.1449996381584022,
|
| 1025 |
+
"sacrebleu_ci_high": 0.19615430198730424
|
| 1026 |
+
},
|
| 1027 |
+
"mt_flores_101_eng_ron": {
|
| 1028 |
+
"num_of_instances": 66,
|
| 1029 |
+
"counts": [
|
| 1030 |
+
1326,
|
| 1031 |
+
841,
|
| 1032 |
+
567,
|
| 1033 |
+
389
|
| 1034 |
+
],
|
| 1035 |
+
"totals": [
|
| 1036 |
+
4837,
|
| 1037 |
+
4771,
|
| 1038 |
+
4705,
|
| 1039 |
+
4639
|
| 1040 |
+
],
|
| 1041 |
+
"precisions": [
|
| 1042 |
+
0.2741368616911309,
|
| 1043 |
+
0.17627331796269124,
|
| 1044 |
+
0.12051009564293304,
|
| 1045 |
+
0.0838542789394266
|
| 1046 |
+
],
|
| 1047 |
+
"bp": 1.0,
|
| 1048 |
+
"sys_len": 4837,
|
| 1049 |
+
"ref_len": 1949,
|
| 1050 |
+
"sacrebleu": 0.14865368140396532,
|
| 1051 |
+
"score": 0.14865368140396532,
|
| 1052 |
+
"score_name": "sacrebleu",
|
| 1053 |
+
"score_ci_low": 0.12285617051153769,
|
| 1054 |
+
"score_ci_high": 0.20722793764019112,
|
| 1055 |
+
"sacrebleu_ci_low": 0.12285617051153769,
|
| 1056 |
+
"sacrebleu_ci_high": 0.20722793764019112
|
| 1057 |
+
},
|
| 1058 |
+
"mt_flores_101_eng_spa": {
|
| 1059 |
+
"num_of_instances": 66,
|
| 1060 |
+
"counts": [
|
| 1061 |
+
1381,
|
| 1062 |
+
759,
|
| 1063 |
+
459,
|
| 1064 |
+
281
|
| 1065 |
+
],
|
| 1066 |
+
"totals": [
|
| 1067 |
+
6405,
|
| 1068 |
+
6339,
|
| 1069 |
+
6273,
|
| 1070 |
+
6207
|
| 1071 |
+
],
|
| 1072 |
+
"precisions": [
|
| 1073 |
+
0.21561280249804843,
|
| 1074 |
+
0.11973497397065783,
|
| 1075 |
+
0.07317073170731708,
|
| 1076 |
+
0.04527146769776059
|
| 1077 |
+
],
|
| 1078 |
+
"bp": 1.0,
|
| 1079 |
+
"sys_len": 6405,
|
| 1080 |
+
"ref_len": 2098,
|
| 1081 |
+
"sacrebleu": 0.09616441307249027,
|
| 1082 |
+
"score": 0.09616441307249027,
|
| 1083 |
+
"score_name": "sacrebleu",
|
| 1084 |
+
"score_ci_low": 0.07930542472032699,
|
| 1085 |
+
"score_ci_high": 0.11825351259228321,
|
| 1086 |
+
"sacrebleu_ci_low": 0.07930542472032699,
|
| 1087 |
+
"sacrebleu_ci_high": 0.11825351259228321
|
| 1088 |
+
},
|
| 1089 |
+
"mt_flores_101_fra_eng": {
|
| 1090 |
+
"num_of_instances": 66,
|
| 1091 |
+
"counts": [
|
| 1092 |
+
1391,
|
| 1093 |
+
930,
|
| 1094 |
+
644,
|
| 1095 |
+
448
|
| 1096 |
+
],
|
| 1097 |
+
"totals": [
|
| 1098 |
+
5836,
|
| 1099 |
+
5770,
|
| 1100 |
+
5704,
|
| 1101 |
+
5638
|
| 1102 |
+
],
|
| 1103 |
+
"precisions": [
|
| 1104 |
+
0.23834818368745714,
|
| 1105 |
+
0.1611785095320624,
|
| 1106 |
+
0.11290322580645162,
|
| 1107 |
+
0.07946080170273147
|
| 1108 |
+
],
|
| 1109 |
+
"bp": 1.0,
|
| 1110 |
+
"sys_len": 5836,
|
| 1111 |
+
"ref_len": 1734,
|
| 1112 |
+
"sacrebleu": 0.13625252798661136,
|
| 1113 |
+
"score": 0.13625252798661136,
|
| 1114 |
+
"score_name": "sacrebleu",
|
| 1115 |
+
"score_ci_low": 0.11732338312550004,
|
| 1116 |
+
"score_ci_high": 0.15953611321490402,
|
| 1117 |
+
"sacrebleu_ci_low": 0.11732338312550004,
|
| 1118 |
+
"sacrebleu_ci_high": 0.15953611321490402
|
| 1119 |
+
},
|
| 1120 |
+
"mt_flores_101_jpn_eng": {
|
| 1121 |
+
"num_of_instances": 66,
|
| 1122 |
+
"counts": [
|
| 1123 |
+
1250,
|
| 1124 |
+
617,
|
| 1125 |
+
343,
|
| 1126 |
+
187
|
| 1127 |
+
],
|
| 1128 |
+
"totals": [
|
| 1129 |
+
6443,
|
| 1130 |
+
6377,
|
| 1131 |
+
6311,
|
| 1132 |
+
6245
|
| 1133 |
+
],
|
| 1134 |
+
"precisions": [
|
| 1135 |
+
0.1940090020176936,
|
| 1136 |
+
0.09675395954210445,
|
| 1137 |
+
0.054349548407542386,
|
| 1138 |
+
0.029943955164131304
|
| 1139 |
+
],
|
| 1140 |
+
"bp": 1.0,
|
| 1141 |
+
"sys_len": 6443,
|
| 1142 |
+
"ref_len": 1734,
|
| 1143 |
+
"sacrebleu": 0.07434451508550546,
|
| 1144 |
+
"score": 0.07434451508550546,
|
| 1145 |
+
"score_name": "sacrebleu",
|
| 1146 |
+
"score_ci_low": 0.0644058696100964,
|
| 1147 |
+
"score_ci_high": 0.08685819903559298,
|
| 1148 |
+
"sacrebleu_ci_low": 0.0644058696100964,
|
| 1149 |
+
"sacrebleu_ci_high": 0.08685819903559298
|
| 1150 |
+
},
|
| 1151 |
+
"mt_flores_101_kor_eng": {
|
| 1152 |
+
"num_of_instances": 66,
|
| 1153 |
+
"counts": [
|
| 1154 |
+
1162,
|
| 1155 |
+
546,
|
| 1156 |
+
295,
|
| 1157 |
+
168
|
| 1158 |
+
],
|
| 1159 |
+
"totals": [
|
| 1160 |
+
5914,
|
| 1161 |
+
5848,
|
| 1162 |
+
5782,
|
| 1163 |
+
5716
|
| 1164 |
+
],
|
| 1165 |
+
"precisions": [
|
| 1166 |
+
0.19648292188028407,
|
| 1167 |
+
0.09336525307797537,
|
| 1168 |
+
0.05102040816326531,
|
| 1169 |
+
0.02939118264520644
|
| 1170 |
+
],
|
| 1171 |
+
"bp": 1.0,
|
| 1172 |
+
"sys_len": 5914,
|
| 1173 |
+
"ref_len": 1734,
|
| 1174 |
+
"sacrebleu": 0.07242154227292893,
|
| 1175 |
+
"score": 0.07242154227292893,
|
| 1176 |
+
"score_name": "sacrebleu",
|
| 1177 |
+
"score_ci_low": 0.06291349695976677,
|
| 1178 |
+
"score_ci_high": 0.08531939881489306,
|
| 1179 |
+
"sacrebleu_ci_low": 0.06291349695976677,
|
| 1180 |
+
"sacrebleu_ci_high": 0.08531939881489306
|
| 1181 |
+
},
|
| 1182 |
+
"mt_flores_101_por_eng": {
|
| 1183 |
+
"num_of_instances": 66,
|
| 1184 |
+
"counts": [
|
| 1185 |
+
1417,
|
| 1186 |
+
974,
|
| 1187 |
+
723,
|
| 1188 |
+
547
|
| 1189 |
+
],
|
| 1190 |
+
"totals": [
|
| 1191 |
+
6392,
|
| 1192 |
+
6326,
|
| 1193 |
+
6260,
|
| 1194 |
+
6194
|
| 1195 |
+
],
|
| 1196 |
+
"precisions": [
|
| 1197 |
+
0.22168335419274093,
|
| 1198 |
+
0.15396775213404995,
|
| 1199 |
+
0.11549520766773164,
|
| 1200 |
+
0.08831126896997095
|
| 1201 |
+
],
|
| 1202 |
+
"bp": 1.0,
|
| 1203 |
+
"sys_len": 6392,
|
| 1204 |
+
"ref_len": 1734,
|
| 1205 |
+
"sacrebleu": 0.13659529350611513,
|
| 1206 |
+
"score": 0.13659529350611513,
|
| 1207 |
+
"score_name": "sacrebleu",
|
| 1208 |
+
"score_ci_low": 0.11528113587288888,
|
| 1209 |
+
"score_ci_high": 0.15836557569065857,
|
| 1210 |
+
"sacrebleu_ci_low": 0.11528113587288888,
|
| 1211 |
+
"sacrebleu_ci_high": 0.15836557569065857
|
| 1212 |
+
},
|
| 1213 |
+
"mt_flores_101_ron_eng": {
|
| 1214 |
+
"num_of_instances": 66,
|
| 1215 |
+
"counts": [
|
| 1216 |
+
1395,
|
| 1217 |
+
948,
|
| 1218 |
+
686,
|
| 1219 |
+
506
|
| 1220 |
+
],
|
| 1221 |
+
"totals": [
|
| 1222 |
+
5680,
|
| 1223 |
+
5614,
|
| 1224 |
+
5548,
|
| 1225 |
+
5482
|
| 1226 |
+
],
|
| 1227 |
+
"precisions": [
|
| 1228 |
+
0.24559859154929575,
|
| 1229 |
+
0.16886355539722125,
|
| 1230 |
+
0.12364816149963952,
|
| 1231 |
+
0.09230207953301715
|
| 1232 |
+
],
|
| 1233 |
+
"bp": 1.0,
|
| 1234 |
+
"sys_len": 5680,
|
| 1235 |
+
"ref_len": 1734,
|
| 1236 |
+
"sacrebleu": 0.14749939449155924,
|
| 1237 |
+
"score": 0.14749939449155924,
|
| 1238 |
+
"score_name": "sacrebleu",
|
| 1239 |
+
"score_ci_low": 0.1292139766827155,
|
| 1240 |
+
"score_ci_high": 0.1819243976706418,
|
| 1241 |
+
"sacrebleu_ci_low": 0.1292139766827155,
|
| 1242 |
+
"sacrebleu_ci_high": 0.1819243976706418
|
| 1243 |
+
},
|
| 1244 |
+
"mt_flores_101_spa_eng": {
|
| 1245 |
+
"num_of_instances": 66,
|
| 1246 |
+
"counts": [
|
| 1247 |
+
1300,
|
| 1248 |
+
712,
|
| 1249 |
+
450,
|
| 1250 |
+
290
|
| 1251 |
+
],
|
| 1252 |
+
"totals": [
|
| 1253 |
+
6166,
|
| 1254 |
+
6100,
|
| 1255 |
+
6034,
|
| 1256 |
+
5968
|
| 1257 |
+
],
|
| 1258 |
+
"precisions": [
|
| 1259 |
+
0.21083360363282516,
|
| 1260 |
+
0.11672131147540984,
|
| 1261 |
+
0.07457739476300962,
|
| 1262 |
+
0.04859249329758713
|
| 1263 |
+
],
|
| 1264 |
+
"bp": 1.0,
|
| 1265 |
+
"sys_len": 6166,
|
| 1266 |
+
"ref_len": 1734,
|
| 1267 |
+
"sacrebleu": 0.09717769541368253,
|
| 1268 |
+
"score": 0.09717769541368253,
|
| 1269 |
+
"score_name": "sacrebleu",
|
| 1270 |
+
"score_ci_low": 0.08585452687395417,
|
| 1271 |
+
"score_ci_high": 0.11046953561336194,
|
| 1272 |
+
"sacrebleu_ci_low": 0.08585452687395417,
|
| 1273 |
+
"sacrebleu_ci_high": 0.11046953561336194
|
| 1274 |
+
},
|
| 1275 |
+
"score": 0.11283376706964308,
|
| 1276 |
+
"score_name": "subsets_mean",
|
| 1277 |
+
"num_of_instances": 990
|
| 1278 |
+
},
|
| 1279 |
+
"score": 0.4523639482757903,
|
| 1280 |
+
"score_name": "subsets_mean",
|
| 1281 |
+
"num_of_instances": 12472
|
| 1282 |
+
}
|
| 1283 |
+
}
|