app.py

import os
import gradio as gr
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from io import BytesIO
from PIL import Image
from datasets.exceptions import DatasetNotFoundError

from src.dataloading import get_leaderboard_models_cached, get_leaderboard_datasets
from src.similarity import load_data_and_compute_similarities

# Set matplotlib backend for non-GUI environments
plt.switch_backend('Agg')


def create_heatmap(selected_models, selected_dataset, selected_metric):
    if not selected_models or not selected_dataset:
        return None
    
    # Sort models and get short names
    similarities = load_data_and_compute_similarities(selected_models, selected_dataset, selected_metric)

    # Check if similarity matrix contains NaN rows
    failed_models = []
    for i in range(len(similarities)):
        if np.isnan(similarities[i]).all():
            failed_models.append(selected_models[i])

    if failed_models:
        gr.Warning(f"Failed to load data for models: {', '.join(failed_models)}")

    # Create figure and heatmap using seaborn
    plt.figure(figsize=(8, 6))
    ax = sns.heatmap(
        similarities,
        annot=True,
        fmt=".2f",
        cmap="viridis",
        vmin=0,
        vmax=1,
        xticklabels=selected_models,
        yticklabels=selected_models
    )
    
    # Customize plot
    plt.title(f"{selected_metric} for {selected_dataset}", fontsize=16)
    plt.xlabel("Models", fontsize=14)
    plt.ylabel("Models", fontsize=14)
    plt.xticks(rotation=45, ha='right')
    plt.yticks(rotation=0)
    plt.tight_layout()

    # Save to buffer
    buf = BytesIO()
    plt.savefig(buf, format="png", dpi=100, bbox_inches="tight")
    plt.close()
    
    # Convert to PIL Image
    buf.seek(0)
    img = Image.open(buf).convert("RGB")
    return img

def validate_inputs(selected_models, selected_dataset):
    if not selected_models:
        raise gr.Error("Please select at least one model!")
    if not selected_dataset:
        raise gr.Error("Please select a dataset!")
    

def update_datasets_based_on_models(selected_models, current_dataset):
    try:
        available_datasets = get_leaderboard_datasets(selected_models) if selected_models else []
        if current_dataset in available_datasets:
            valid_dataset = current_dataset
        elif "mmlu_pro" in available_datasets:
            valid_dataset = "mmlu_pro"
        else:
            valid_dataset = None
        return gr.update(
            choices=available_datasets,
            value=valid_dataset
        )
    except DatasetNotFoundError as e:
        # Extract model name from error message
        model_name = e.args[0].split("'")[1]
        model_name = model_name.split("/")[-1].replace("__", "/").replace("_details", "")
        
        # Display a shorter warning
        gr.Warning(f"Data for '{model_name}' is gated or unavailable.")
        return gr.update(choices=[], value=None)

links_markdown = """
[📄 Paper](https://arxiv.org/abs/2502.04313)   |  
[☯ Homepage](https://model-similarity.github.io/)   |  
[🐱 Code](https://github.com/model-similarity/lm-similarity)   |  
[🐍 pip install lm-sim](https://pypi.org/project/lm-sim/)   |  
[🤗 Data](https://huggingface.co/datasets/bethgelab/lm-similarity)
"""

model_init = ["HuggingFaceTB/SmolLM2-1.7B-Instruct", "meta-llama/Llama-3.1-8B-Instruct", "microsoft/phi-4", "google/gemma-2-27b-it", "Qwen/Qwen2.5-32B-Instruct", "meta-llama/Llama-3.3-70B-Instruct"]
dataset_init = "mmlu_pro"
metric_init = "CAPA"


# Create Gradio interface
with gr.Blocks(title="LLM Similarity Analyzer") as demo:
    gr.Markdown("# Model Similarity Comparison Tool")
    gr.Markdown(links_markdown)
    gr.Markdown('Demo for the recent publication ["Great Models Think Alike and this Undermines AI Oversight"](https://huggingface.co/papers/2502.04313).')

    with gr.Row():
        dataset_dropdown = gr.Dropdown(
            choices=get_leaderboard_datasets(model_init),
            label="Select Dataset",
            value=dataset_init,
            filterable=True,
            interactive=True,
            allow_custom_value=False,
            info="Open LLM Leaderboard v2 benchmark datasets"
        )
        metric_dropdown = gr.Dropdown(
            choices=["CAPA", "CAPA (det.)", "Error Consistency"],
            label="Select Metric",
            value=metric_init,
            info="Select a similarity metric to compute"
        )

    model_dropdown = gr.Dropdown(
        choices=get_leaderboard_models_cached(),
        label="Select Models",
        value=model_init,
        multiselect=True,
        filterable=True,
        allow_custom_value=False,
        info="Search and select multiple models"
    )

    model_dropdown.change(
        fn=update_datasets_based_on_models,
        inputs=[model_dropdown, dataset_dropdown],
        outputs=dataset_dropdown
    )
    
    generate_btn = gr.Button("Generate Heatmap", variant="primary")
    heatmap = gr.Image(value=create_heatmap(model_init, dataset_init, metric_init), label="Similarity Heatmap", visible=True)

    generate_btn.click(
        fn=validate_inputs,
        inputs=[model_dropdown, dataset_dropdown],
        queue=False
    ).then(
        fn=create_heatmap,
        inputs=[model_dropdown, dataset_dropdown, metric_dropdown],
        outputs=heatmap
    )
    
    gr.Markdown("\* Self-similarity is only 1.0 for the probabilistic Kappa_p metric if the model predicts a single option with 100% confidence for each question.")

    clear_btn = gr.Button("Clear Selection")
    clear_btn.click(
        lambda: [[], None, None],
        outputs=[model_dropdown, dataset_dropdown, heatmap]
    )

    gr.Markdown("## Information")
    gr.Markdown("""We propose Chance Adjusted Probabilistic Agreement ($\operatorname\{CAPA\}$, or $\kappa_p$), a novel metric \
for model similarity which adjusts for chance agreement due to accuracy. Using CAPA, we find: (1) LLM-as-a-judge scores are \
biased towards more similar models controlling for the model's capability. (2) Gain from training strong models on annotations \
of weak supervisors (weak-to-strong generalization) is higher when the two models are more different. (3) Concerningly, model \
errors are getting more correlated as capabilities increase.""")
    with gr.Row():
        gr.Image(value="data/table_capa.png", label="Comparison of different similarity metrics for multiple-choice questions", interactive=False, scale=1)
    gr.Markdown("""
- **Datasets**: [Open LLM Leaderboard v2](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/) benchmark datasets \n
    - Some datasets are not multiple-choice - for these, the metrics are not applicable. \n
- **Models**: Open LLM Leaderboard models \n
    - Every model evaluation is gated on Hugging Face and access has to be requested. \n
    - We requested access for the most popular models, but some may be missing. \n
    - Notably, loading data is not possible for some meta-llama and gemma models.
- **Metrics**: CAPA (probabilistic), CAPA (deterministic), Error Consistency""")

if __name__ == "__main__":
    demo.launch(ssr_mode=False)