The Context Course

You are coordinating NanoChat post-training experiments in this project using Pi and pi-subagents.

Arguments:

• campaign: $1 if provided, otherwise nanochat post-training improvements
• gpu slots: $2 if provided, otherwise 1
• max ideas: $3 if provided, otherwise 3

Read:

• AGENTS.md
• README.md
• program.md
• docs/pi-subagents-guide.md
• research/notes.md
• research/results.tsv
• train.py
• prepare.py
• evaluate.py
• model.py
• scripts/hf_job.py

Use Pi project agents:

• planner for up to the requested max ideas for the requested campaign
• reviewer for read-only rule and comparability checks
• researcher for paper, nanochat, or PostTrainBench-derived hypotheses
• reporter for Hugging Face Jobs status and logs
• memory-keeper for durable notes in this checkout
• experiment-worker for one coherent post-training method change

Reject ideas that modify evaluate.py, alter the submitted NanoChat architecture, train on eval examples, or mix unrelated method changes.

Do not allow more active managed experiment jobs than the requested GPU slots. For implementation work, ask experiment-worker to make one coherent change, run a local smoke test, then launch the managed Hugging Face Jobs benchmark if the smoke test passes.

Keep experiments comparable:

• work from this post-training project root, not ../pre-training
• prefer editing train.py only unless a small supporting file is justified
• keep model.py architecture-compatible with the fixed evaluator
• do not modify evaluate.py or src/eval/tasks/*/evaluate.py
• do not train on eval examples
• save the best checkpoint to final_model/
• run the fixed evaluator before reporting a score
• persist managed results under the mounted POSTTRAIN_HF_BUCKET
• update research/notes.md and research/results.tsv after completed runs

Suggested managed commands:

uv run scripts/hf_job.py launch --mode experiment
uv run scripts/hf_job.py logs <JOB_ID> --follow --output /tmp/posttrain-run.log