Mini Parameter Golf

Mini Parameter Golf is a two-hour workshop version of OpenAI Parameter Golf. The task is to improve solution.py for lower val_bpb. The original challenge is designed for rigorous evaluation; Mini PG is designed for educational iteration with lower runtime and cost. In one saved 20-step Weco run, steps averaged about 2.3 minutes end to end.

For allowed edits and scoring constraints, read docs/rules.md.

1. Set Up

If you do not have uv, install it from the uv install docs.

git clone https://github.com/WecoAI/mini-pg.git
cd mini-pg
uv sync
uv run weco credits balance
uv run modal setup

If uv run weco credits balance reports an auth error, run uv run weco login. For Modal authentication and troubleshooting, see docs/modal.md.

If you are using Claude Code to start the Weco run, this dashboard-bridged session is required before section 2. Start it and do not start a separate direct run from the original shell.

uv run weco start claude --headless \
  -p "Read README.md and docs/rules.md, then run ./scripts/weco.sh to optimize solution.py for lower val_bpb. Do not run uv run weco run directly. Monitor the run and report new bests; don't block."

Use the Weco dashboard to watch and steer the bridged session.

2. Run with Weco

Run the default Mini PG search:

uv run weco run \
  --source solution.py \
  --eval-command "./scripts/weco_eval.sh solution.py" \
  --metric val_bpb \
  --goal minimize \
  --steps 20 \
  --eval-timeout 210 \
  --save-logs \
  --additional-instructions instructions/weco_instructions.md \
  --output plain \
  --no-open

Weco prints a run link or run ID. Use it to watch progress and collect results. For monitoring, steering, and reporting details, see the appendix in docs/appendix.md. For general Weco usage, see the Weco docs, the Weco CLI repository, and the Weco examples.

Leaderboard

Lower val_bpb is better.

Record ID	val_bpb	runtime_s	artifact_bytes	Method	Record
20260624-1310-baseline	3.983848	39.016	1,265,674	Baseline 4-layer byte Transformer, learned positions, default quantization, 80 train steps	record/20260624-1310-baseline

To submit a leaderboard PR, add a record under record/ using the record format, then update this table.

Acknowledgements

Mini Parameter Golf is inspired by OpenAI's official Parameter Golf challenge. Thanks to the OpenAI Parameter Golf authors and maintainers, including @cocohearts, @0hq, and @valerio-oai.

Citation

If you use Mini Parameter Golf in a workshop, project, or write-up, please cite this repository:

@software{wecoai_mini_pg_2026,
  author = {{WecoAI}},
  title = {Mini Parameter Golf},
  year = {2026},
  url = {https://github.com/WecoAI/mini-pg}
}