autoku

Autonomous NxN Sudoku solver research. Trains a GPT to solve Sudoku puzzles, inspired by karpathy/autoresearch.

Quick Start

uv sync
uv run prepare.py          # sanity check
uv run train.py             # train for 15 minutes

Ralph Mode

To start an autonomous experiment loop using Claude Code:

/ralph-loop

You are an autonomous ML researcher. Read program.md for full instructions.

Setup:
1. Read program.md, prepare.py, and train.py to understand the codebase
2. Create branch: git checkout -b autoku/<today's date, e.g. apr3>
3. Create results.tsv with header: commit	val_avg_tokens	val_solve_rate	val_acc	status	description
4. Run baseline: uv run train.py > run.log 2>&1
5. Log baseline to results.tsv

Then enter the experiment loop from program.md:
- Modify train.py with an idea (architecture, hyperparams, optimizer, model size)
- git commit
- uv run train.py > run.log 2>&1
- grep "^val_avg_tokens:\|^val_solve_rate:\|^val_acc:" run.log
- If val_avg_tokens decreased: keep. If not: git reset HEAD~1
- Log to results.tsv
- NEVER STOP. Loop until interrupted.

Goal: minimize val_avg_tokens (Las Vegas: always correct, minimize average cost).
Start with 4x4, bump to 9x9 when val_solve_rate > 0.95 and val_avg_tokens is near optimal.