| from env import TASK | |
| from utils import run_pipeline, update_examples | |
| import gradio as gr | |
| with gr.Blocks( | |
| title="YourBench Leaderboard", | |
| css="button { margin: 0 10px; padding: 5px 15px; }", | |
| ) as app: | |
| # DISPLAY TABLE AND ANALYSIS | |
| title = gr.Markdown(f"YourBench auto-Leaderboard for {TASK}") | |
| leaderboard = gr.DataFrame(label="Results", interactive=False) | |
| samples_ix = gr.Number(label="Example Index", value=0, step=1, info="Navigate through different examples") | |
| with gr.Tab("Hardest samples"): | |
| hard_samples = gr.HTML() | |
| with gr.Tab("Easiest samples"): | |
| easy_samples = gr.HTML() | |
| with gr.Tab("All samples"): | |
| all_samples = gr.HTML() | |
| samples_ix.change(update_examples, samples_ix, [easy_samples, hard_samples, all_samples]) | |
| app.load(run_pipeline, [samples_ix], [leaderboard, easy_samples, hard_samples, all_samples]) | |
| app.launch() | |