Spaces:
Runtime error
Runtime error
Commit
·
98ced8b
1
Parent(s):
1af55c6
add: train_binary_classifier
Browse files
guardrails_genie/train_classifier.py
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import evaluate
|
| 2 |
+
import numpy as np
|
| 3 |
+
import wandb
|
| 4 |
+
from datasets import load_dataset
|
| 5 |
+
from transformers import (
|
| 6 |
+
AutoModelForSequenceClassification,
|
| 7 |
+
AutoTokenizer,
|
| 8 |
+
DataCollatorWithPadding,
|
| 9 |
+
Trainer,
|
| 10 |
+
TrainingArguments,
|
| 11 |
+
)
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def train_binary_classifier(
|
| 15 |
+
project_name: str,
|
| 16 |
+
entity_name: str,
|
| 17 |
+
dataset_repo: str = "geekyrakshit/prompt-injection-dataset",
|
| 18 |
+
model_name: str = "distilbert/distilbert-base-uncased",
|
| 19 |
+
learning_rate: float = 2e-5,
|
| 20 |
+
batch_size: int = 16,
|
| 21 |
+
num_epochs: int = 2,
|
| 22 |
+
weight_decay: float = 0.01,
|
| 23 |
+
):
|
| 24 |
+
wandb.init(project=project_name, entity=entity_name)
|
| 25 |
+
dataset = load_dataset(dataset_repo)
|
| 26 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 27 |
+
|
| 28 |
+
def preprocess_function(examples):
|
| 29 |
+
return tokenizer(examples["prompt"], truncation=True)
|
| 30 |
+
|
| 31 |
+
tokenized_datasets = dataset.map(preprocess_function, batched=True)
|
| 32 |
+
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
|
| 33 |
+
accuracy = evaluate.load("accuracy")
|
| 34 |
+
|
| 35 |
+
def compute_metrics(eval_pred):
|
| 36 |
+
predictions, labels = eval_pred
|
| 37 |
+
predictions = np.argmax(predictions, axis=1)
|
| 38 |
+
return accuracy.compute(predictions=predictions, references=labels)
|
| 39 |
+
|
| 40 |
+
id2label = {0: "SAFE", 1: "INJECTION"}
|
| 41 |
+
label2id = {"SAFE": 0, "INJECTION": 1}
|
| 42 |
+
|
| 43 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
| 44 |
+
model_name,
|
| 45 |
+
num_labels=2,
|
| 46 |
+
id2label=id2label,
|
| 47 |
+
label2id=label2id,
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
trainer = Trainer(
|
| 51 |
+
model=model,
|
| 52 |
+
args=TrainingArguments(
|
| 53 |
+
output_dir="binary-classifier",
|
| 54 |
+
learning_rate=learning_rate,
|
| 55 |
+
per_device_train_batch_size=batch_size,
|
| 56 |
+
per_device_eval_batch_size=batch_size,
|
| 57 |
+
num_train_epochs=num_epochs,
|
| 58 |
+
weight_decay=weight_decay,
|
| 59 |
+
eval_strategy="epoch",
|
| 60 |
+
save_strategy="epoch",
|
| 61 |
+
load_best_model_at_end=True,
|
| 62 |
+
push_to_hub=True,
|
| 63 |
+
report_to="wandb",
|
| 64 |
+
),
|
| 65 |
+
train_dataset=tokenized_datasets["train"],
|
| 66 |
+
eval_dataset=tokenized_datasets["test"],
|
| 67 |
+
processing_class=tokenizer,
|
| 68 |
+
data_collator=data_collator,
|
| 69 |
+
compute_metrics=compute_metrics,
|
| 70 |
+
)
|
| 71 |
+
trainer.train()
|