app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
import datasets

# Load your fine-tuned model and tokenizer
model_name = "legacy107/flan-t5-large-bottleneck-adapter-cpgQA-unique"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
model.set_active_adapters("question_answering")
max_length = 512
max_target_length = 128

# Load your dataset
dataset = datasets.load_dataset("minh21/cpgQA-v1.0-unique-context-test-10-percent", split="test[:10]")


# Define your function to generate answers
def generate_answer(question, context):
    # Combine question and context
    input_text = f"question: {question} context: {context}"

    # Tokenize the input text
    input_ids = tokenizer(
        input_text,
        return_tensors="pt",
        padding="max_length",
        truncation=True,
        max_length=max_length,
    ).input_ids

    # Generate the answer
    with torch.no_grad():
        generated_ids = model.generate(input_ids, max_new_tokens=max_target_length)

    # Decode and return the generated answer
    generated_answer = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    return generated_answer


# Define a function to list examples from the dataset
def list_examples():
    examples = []
    for example in dataset:
        context = example["context"]
        question = example["question"]
        examples.append([context, question])
    return examples


# Create a Gradio interface
iface = gr.Interface(
    fn=generate_answer,
    inputs=[
        gr.inputs.Textbox(label="Question"),
        gr.inputs.Textbox(label="Context")
    ],
    outputs=gr.outputs.Textbox(label="Generated Answer"),
    examples=list_examples()
)

# Launch the Gradio interface
iface.launch()