Convert published benchmark leaderboards from Hugging Face datasets, papers, blog posts, GitHub READMEs, tables, or similar sources into local, human-reviewable draft changes for Hugging Face model repositories using `.eval_results/*.yaml`. Use when mapping benchmark model names to exact Hub model repos, checking native benchmark task IDs, extracting scores with provenance, preventing duplicate benchmark submissions, handling existing eval files or open PRs, and preparing a per-model PR plan. Never open, upload, or submit the model PRs; stop at a review bundle so a human can verify every target and value first.
Add and manage evaluation results in Hugging Face model repositories using the new .eval_results/ format. Supports extracting scores from model cards.