app.py

import spaces
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = 'yuntian-deng/gpt2-implicit-cot-multiplication'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

def preprocess(num):
    num = str(num).strip().replace(' ', '')
    reversed_num = ' '.join(num[::-1])
    return reversed_num

def postprocess(raw_output):
    prediction = raw_output.replace(' ', '')[::-1]
    return prediction

@spaces.GPU
def predict_product(num1, num2):
    input_text = f'{preprocess(num1)} * {preprocess(num2)} ='
    inputs = tokenizer(input_text, return_tensors='pt').to('cuda' if torch.cuda.is_available() else 'cpu')
    model.to('cuda' if torch.cuda.is_available() else 'cpu')

    generated_ids = inputs['input_ids']
    outputs = model.generate(generated_ids, max_new_tokens=40, do_sample=False)
    full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
    prediction = postprocess(full_output[len(input_text):])

    try:
        num1_int = int(num1)
        num2_int = int(num2)
        correct_product = str(num1_int * num2_int)
    except ValueError:
        return [], "Invalid input. Could not evaluate correctness."

    # Create the diff for HighlightedText
    diff = []
    max_len = max(len(prediction), len(correct_product))
    for i in range(max_len):
        if i < len(prediction) and i < len(correct_product) and prediction[i] == correct_product[i]:
            diff.append((prediction[i], None))  # No highlight for correct digits
        elif i < len(prediction) and (i >= len(correct_product) or prediction[i] != correct_product[i]):
            diff.append((prediction[i], "+"))  # Highlight incorrect digits in red
        if i < len(correct_product) and (i >= len(prediction) or prediction[i] != correct_product[i]):
            diff.append((correct_product[i], "-"))  # Highlight missing/incorrect digits in green

    result_message = "Correct!" if prediction == correct_product else f"Incorrect! The correct product is {correct_product}."

    return diff, result_message

demo = gr.Interface(
    fn=predict_product,
    inputs=[
        gr.Textbox(label='First Number (up to 12 digits)', value='12345'),
        gr.Textbox(label='Second Number (up to 12 digits)', value='67890'),
    ],
    outputs=[
        gr.HighlightedText(label='Predicted Product with Matching and Unmatching Digits Highlighted', combine_adjacent=True, show_legend=True, color_map={"-": "green", "+": "red"}),
        gr.HTML(label='Result Message')
    ],
    title='GPT2 Direct Multiplication Calculator (Without Using Chain-of-Thought)',
    description='This demo uses GPT2 to directly predict the product of two numbers without using any intermediate reasoning steps. The GPT2 model has been fine-tuned to internalize chain-of-thought reasoning within its hidden states, following our stepwise internalization approach detailed in the paper linked at the bottom of this page.',
    article="""
    - [Paper: From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://arxiv.org/pdf/2405.14838)
    - [Code Repository](https://github.com/da03/Internalize_CoT_Step_by_Step)
    - [Tweet Announcement](https://twitter.com/yuntiandeng/status/1795854740879774036)
    """,
    clear_btn=None,
    submit_btn="Multiply!",
    live=False
)

demo.launch()