app.py

import json
import random

import gradio as gr
from difflib import SequenceMatcher

file_path = "dataset.jsonl"
similarity_threshold = 0.85
current_index = 0 

description_text = """
This Space is inspired by [Luis Hunt's](https://www.linkedin.com/posts/louiswhunt_see-below-for-6882-pages-of-mmlu-and-gsm8k-activity-7281011488692047872-fWCE?utm_source=share&utm_medium=member_desktop) post.
He highlights how current top performing models from major vendors are contaminated with benchmark data that is supposed to be used to assess their performance.

This space aims to partially reproduce this work. I chose to look at the contamination of **Qwen/Qwen2.5-14B** by **GSM8K** dataset.
"""


def find_similar_chunks(original, output):
    matcher = SequenceMatcher(None, original, output)
    left = 0
    highlighted_sequence = []
    for _, j, n in matcher.get_matching_blocks():
        if left < j:
            highlighted_sequence.append((output[left:j], None))
        highlighted_sequence.append((output[j:j+n], 1))
        left = j + n
    if j+n < len(output) - 1:
        highlighted_sequence.append((output[j+n:], None))
    
    return highlighted_sequence

with open(file_path, "r") as file:
    examples = [json.loads(line) for line in file if json.loads(line)["similarity_ratio"] > similarity_threshold]

def next_example():
    new_example = random.choice(examples)

    highlighted_output = find_similar_chunks(new_example["original"], new_example["output"])
    return(
        [
            new_example["prompt"],
            new_example["original"], 
            highlighted_output, 
            new_example["similarity_ratio"],
            new_example["seed"]
        ]
    )

with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown(description_text)
        with gr.Column(scale=1):
            pass
    prompt = gr.Textbox(
        label="Prompt",
        interactive=False,
        value=examples[current_index]["prompt"],
    )
    with gr.Row():
        with gr.Column(scale=4):
            original = gr.Textbox(
                label="Original",
                interactive=False,
                value=examples[current_index]["original"],
            )
        with gr.Column(scale=4):
            output = gr.HighlightedText(
                label="Output",
                color_map={"1": "yellow"},
                value=find_similar_chunks(examples[current_index]["original"], 
                                        examples[current_index]["output"]),
            )

        with gr.Row():
            with gr.Column(scale=1):
                similarity = gr.Textbox(
                    label="Similarity ratio",
                    interactive=False,
                    value=examples[current_index]["similarity_ratio"],
                )
            with gr.Column(scale=1):
                seed = gr.Textbox(
                    label="Seed",
                    interactive=False,
                    value=examples[current_index]["seed"],
                )

    next_btn = gr.Button("Anoter example")
    
    next_btn.click(fn=next_example, 
                    outputs=[prompt, original, output, similarity, seed])



demo.launch()