Spaces:
Sleeping
Sleeping
| import gradio as gr | |
| import pandas as pd | |
| def load_and_process_data(file_path): | |
| # Load the leaderboard data | |
| df = pd.read_pickle(file_path) | |
| # Group by 'lmm' and 'question' to calculate mean accuracy | |
| accuracy_df = ( | |
| df.groupby(["lmm", "question"])["accepted_by_judge"].mean().reset_index() | |
| ) | |
| accuracy_df = accuracy_df.rename(columns={"accepted_by_judge": "accuracy"}) | |
| accuracy_df["accuracy"] = (accuracy_df["accuracy"] * 100).round(1) | |
| # Group by 'lmm' to calculate the count of images | |
| image_count_df = df.groupby("lmm")["image"].nunique().reset_index() | |
| image_count_df = image_count_df.rename(columns={"image": "Total Images"}) | |
| return accuracy_df, image_count_df | |
| def expand_and_format_df(accuracy_df, image_count_df): | |
| # Pivot and format the accuracy dataframe | |
| expanded_df = accuracy_df.pivot(index="lmm", columns="question", values="accuracy") | |
| expanded_df["Average"] = expanded_df.mean(axis=1).round(1) | |
| expanded_df = expanded_df.sort_values(by="Average", ascending=False).reset_index() | |
| expanded_df.columns.name = None | |
| # Merge the 'total_images' column | |
| final_df = pd.merge(expanded_df, image_count_df, on="lmm") | |
| return final_df.rename(columns={"lmm": "Model"}) | |
| def map_model_names(df, name_dict): | |
| # Map model names using the provided dictionary | |
| df["Model"] = df["Model"].map(name_dict) | |
| return df | |
| # Dictionary for renaming models | |
| name_dict = { | |
| "gpt4v": "GPT-4V(ision)", | |
| "llava": "LLaVA-1.5-13B", | |
| "llava-7b": "LLaVA-1.5-7B", | |
| "Long-SPHINX": "Long-SPHINX", | |
| "SPHINX": "SPHINX", | |
| "OtterHD": "OtterHD", | |
| "minigpt4v2": "MiniGPT4v2", | |
| "InstructBLIP-13B": "InstructBLIP-13B", | |
| "InstructBLIP": "InstructBLIP-7B", | |
| "qwen": "Qwen-VL-Chat", | |
| "fuyu-8b": "Fuyu-8B", | |
| } | |
| # Processing steps | |
| accuracy_df, image_count_df = load_and_process_data("raw_outputs.pkl") | |
| final_df = expand_and_format_df(accuracy_df, image_count_df) | |
| final_df = map_model_names(final_df, name_dict) | |
| # Gradio interface | |
| with gr.Blocks() as demo: | |
| gr.Markdown("# GlitchBench Leaderboard") | |
| gr.HTML(""" | |
| <div style="text-align: center;"> | |
| <a href="https://glitchbench.github.io/" target="_blank">Visit the GlitchBench Project Homepage</a> | |
| </div> | |
| """) | |
| with gr.Row(): | |
| gr.Dataframe(final_df) | |
| gr.Markdown("# How to Submit Your Model") | |
| gr.Markdown( | |
| """ | |
| We warmly invite you to submit your model's responses for inclusion in our leaderboard. To participate, please email a CSV file containing your model's responses to [[email protected]](mailto:[email protected]). Successful submissions will be evaluated and will be featured on our leaderboard. | |
| **CSV File Format:** | |
| ``` | |
| ImageId,Question,Output | |
| ``` | |
| - `ImageId`: The ID of the image. (If unsure how to find the Image ID, please [add instructions or a link here]). | |
| - `Question`: The prompt used to query the model. We currently support only two question types: | |
| 1. "What is unusual about this image?" | |
| 2. "What is wrong with this image?" | |
| - `Output`: The response generated by your model. | |
| Ensure that your submissions follow these guidelines for a smooth evaluation process. We look forward to seeing your innovative models in action! | |
| """ | |
| ) | |
| demo.launch() | |