| name | literature-to-recipe |
| description | Turn papers and docs into a concrete training or evaluation recipe (datasets, metrics, hyperparameters, fine-tuning code). |
| always | true |
Literature to recipe
Goal: extract actionable training/eval recipes from sources, not summaries of abstracts.
Workflow
- Anchor — identify 1–3 anchor papers or official docs for the task (use
arxiv_search, arxiv_fetch, web_search, web_fetch, hf_hub_file_fetch for Hub configs).
- Methodology — read sections on data, preprocessing, model, optimization, and evaluation metrics (not the abstract alone).
- Extract — write down: dataset name/version, splits, preprocessing, loss, batching assumptions, LR schedule, epochs/steps, hardware used, and reported metric with the exact benchmark definition.
- Map to workspace — translate into files or
execution_* steps that fit this sandbox (paths, env vars, pinned package versions).
- Validate data — inspect real files in the repo (
read_file, glob_files, search_text); never assume column names.
- Traceability — cite arXiv IDs or URLs for every non-obvious choice.
Output format for the user:
- Sources (bullets with links or ids)
- Recipe (numbered steps)
- Risks / unknowns (what was not stated clearly in the paper)