Program

This repo is a self-contained autoresearch-style benchmark checkout with a local promoted master and a git-tracked run ledger.

Core rules:

• use uv
• edit train.py only unless explicitly authorized otherwise
• never edit prepare.py
• refresh from the current local promoted master before a fresh experiment: uv run scripts/refresh_master.py --fetch-dag
• treat train_orig.py, research/live/master.json, and research/results.tsv as the benchmark source of truth
• run exactly one managed Hugging Face Jobs experiment per hypothesis
• record every completed run with uv run scripts/submit_patch.py --comment "..."
• promotion is local: scripts/submit_patch.py updates train_orig.py and the live master snapshots only when the observed val_bpb beats current master

Primary workflow:

1. uv sync
2. uv run scripts/refresh_master.py --fetch-dag
3. edit train.py
4. uv run scripts/hf_job.py preflight
5. uv run scripts/hf_job.py launch --mode experiment
6. uv run scripts/hf_job.py logs <JOB_ID> --follow --output /tmp/autolab-run.log
7. uv run scripts/submit_patch.py --comment "..."

Upstream compatibility:

• uv run scripts/sync_upstream.py --check compares the upstream-tracked files against karpathy/autoresearch
• uv run scripts/sync_upstream.py --apply copies those upstream files into this repo without touching the local results ledger