| """Live PostTrain Arena leaderboard backed by a Hub dataset.""" |
|
|
| import json |
|
|
| import gradio as gr |
| from huggingface_hub import hf_hub_download |
|
|
|
|
| LEADERBOARD_REPO = 'benchflow/posttrainarena-leaderboard' |
|
|
|
|
| def load_rows(): |
| path = hf_hub_download( |
| LEADERBOARD_REPO, |
| "leaderboard.json", |
| repo_type="dataset", |
| force_download=True, |
| ) |
| records = json.loads(open(path, encoding="utf-8").read()) |
| rows = [] |
| rank = 0 |
| for record in records: |
| if record.get("status") == "succeeded": |
| rank += 1 |
| shown_rank = rank |
| else: |
| shown_rank = None |
| rows.append( |
| [ |
| shown_rank, |
| record.get("submission_id"), |
| record.get("status"), |
| record.get("baseline_score"), |
| record.get("score_after_posttrain"), |
| record.get("delta_score"), |
| record.get("run_id"), |
| record.get("artifact_url"), |
| ] |
| ) |
| return rows |
|
|
|
|
| with gr.Blocks(title="PostTrain Arena Leaderboard") as demo: |
| gr.Markdown("# PostTrain Arena Leaderboard") |
| gr.Markdown( |
| "Continuously updated results from pinned Hugging Face Jobs. " |
| "Scores link to immutable run artifacts." |
| ) |
| table = gr.Dataframe( |
| headers=[ |
| "Rank", |
| "Submission", |
| "Status", |
| "Baseline", |
| "Final", |
| "Delta", |
| "Run", |
| "Artifacts", |
| ], |
| value=load_rows, |
| interactive=False, |
| ) |
| refresh = gr.Button("Refresh") |
| refresh.click(load_rows, outputs=table) |
| timer = gr.Timer(60) |
| timer.tick(load_rows, outputs=table) |
|
|
|
|
| if __name__ == "__main__": |
| demo.launch() |
|
|
