Update app.py

app.py

@@ -11,28 +11,38 @@ device = 'cpu' # if you have a GPU
 tokenizer = T5Tokenizer.from_pretrained('stanfordnlp/SteamSHP-flan-t5-large')
 model = T5ForConditionalGeneration.from_pretrained('stanfordnlp/SteamSHP-flan-t5-large').to(device)
 
+model_list = [
+ 'google/flan-t5-xxl',
+'bigscience/bloomz-7b1',
+'facebook/opt-iml-max-30b',
+'allenai/tk-instruct-11b-def-pos']
+
 HF_TOKEN = os.getenv("HF_TOKEN")
 
 OUTPUTS_DATASET = "HuggingFaceH4/instruction-pilot-outputs-filtered"
 
 ds = load_dataset(OUTPUTS_DATASET, split="train", use_auth_token=HF_TOKEN)
 
-def process():
+def process(model_A, model_B):
     sample_ds = ds.shuffle().select(range(1))
     sample = sample_ds[0]
     prompt = sample["prompt"]
     
     df = pd.DataFrame.from_records(sample["filtered_outputs"])
+    response_A = df[df['model']==model_A]
+    response_B = df[df['model']==model_B]
+    
         
     input_text = "POST: "+ prompt+ "\n\n RESPONSE A: Lime juice, and zest, then freeze in small quantities.\n\n RESPONSE B: Lime marmalade lol\n\n Which response is better? RESPONSE"
     x = tokenizer([input_text], return_tensors='pt').input_ids.to(device)
     y = model.generate(x, max_new_tokens=1)
     prefered = tokenizer.batch_decode(y, skip_special_tokens=True)[0] 
-    return prefered, df
+    return prefered, response_A
 
 title = "Compare Instruction Models to see which one is more helpful"
 interface = gr.Interface(fn=process, 
-                     inputs=[],
+                     inputs=[gr.Dropdown(choices=model_list, value=model_list[0], label='Model A'),
+                            gr.Dropdown(choices=model_list, value=model_list[1], label='Model B')],
                      outputs=[
                               gr.Textbox(label = "Preferred Option"),
                               gr.DataFrame(label = "Model Responses")