Update app.py

app.py

@@ -1,5 +1,5 @@
 import gradio as gr
-from gradio.components import Textbox
+from gradio.components import Textbox, Checkbox
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, T5ForConditionalGeneration
 from peft import PeftModel
 import torch
@@ -19,6 +19,7 @@ cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
 model_name = "google/flan-t5-large"
 peft_name = "legacy107/flan-t5-large-ia3-covidqa"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
+pretrained_model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large")
 model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large")
 model = PeftModel.from_pretrained(model, peft_name)
 
@@ -32,7 +33,7 @@ max_target_length = 200
 # Load your dataset
 dataset = datasets.load_dataset("minh21/COVID-QA-Chunk-64-testset-biencoder-data-90_10", split="train")
 dataset = dataset.shuffle()
-dataset = dataset.select(range(5))
+dataset = dataset.select(range(10))
 
 # Context chunking
 min_sentences_per_chunk = 3
@@ -138,7 +139,7 @@ def retrieve_context(question, contexts):
 
 
 # Define your function to generate answers
-def generate_answer(question, context, ground):
+def generate_answer(question, context, ground, do_pretrained, do_natural):
     contexts = chunk_splitter(clean_data(context))
     context = retrieve_context(question, contexts)
     
@@ -162,9 +163,18 @@ def generate_answer(question, context, ground):
     generated_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
 
     # Paraphrase answer
-    paraphrased_answer = paraphrase_answer(question, generated_answer)
+    paraphrased_answer = ""
+    if do_natural:
+        paraphrased_answer = paraphrase_answer(question, generated_answer)
 
-    return generated_answer, context, paraphrased_answer
+    # Get pretrained model's answer
+    pretrained_answer = ""
+    if do_pretrained:
+        with torch.no_grad():
+            pretrained_generated_ids = pretrained_model.generate(input_ids=input_ids, max_new_tokens=max_target_length)
+            pretrained_answer = tokenizer.decode(pretrained_generated_ids[0], skip_special_tokens=True)
+
+    return generated_answer, context, paraphrased_answer, pretrained_answer
 
 
 # Define a function to list examples from the dataset
@@ -174,7 +184,7 @@ def list_examples():
         context = example["context"]
         question = example["question"]
         answer = example["answer"]
-        examples.append([question, context, answer])
+        examples.append([question, context, answer, True, True])
     return examples
 
 
@@ -184,14 +194,18 @@ iface = gr.Interface(
     inputs=[
         Textbox(label="Question"),
         Textbox(label="Context"),
-        Textbox(label="Ground truth")
+        Textbox(label="Ground truth"),
+        Checkbox(label="Include pretrained model's result"),
+        Checkbox(label="Include natural answer")
     ],
     outputs=[
         Textbox(label="Generated Answer"),
         Textbox(label="Retrieved Context"),
-        Textbox(label="Natural Answer")
+        Textbox(label="Natural Answer"),
+        Textbox(label="Pretrained Model's Answer")
     ],
-    examples=list_examples()
+    examples=list_examples(),
+    examples_per_page=1,
 )
 
 # Launch the Gradio interface