bicycle-rl

Teaching a small model to draw with code, following https://surya.website/rling-qwen-to-paint-with-code: model writes a sketch → render it headlessly → judge the image → reward → repeat.

The target: a bicycle, side view, painted in p5.brush — see bike/. Bicycles because almost nobody can draw one from memory (Velocipedia), so the failures are legible and the judging is fast.

Three GRPO runs of 320 steps each on Qwen2.5-Coder-7B-Instruct. The base model draws 0 recognisable bicycles in 50; run 3 finished with 117 of 128 judged a bicycle, and still painting. The comparison, and what each run's reward bought, is in results/README.md. Adapters: run1 · run2 · run3.

• AGENTS.md — how we work here. Read it first.
• BITACORA.md — lab notebook, becomes the blog post. Every session appends to it.
• bike/ — the whole thing: prompt, render harness, judge, reward, training loop (its own README)
• bike/train/ — the three-phase loop: sample → score → update, as separate processes
• box/ — rent a GPU, set it up, run, mirror it back (its own README)
• results/ — what survives the boxes: per-run logs, scores, prompts, sample sheets
• runs/ — gitignored local mirror, one directory per run
• bicycles-judged/ — the first drawings a judge called a bicycle
• bitacora-assets/ — images the bitácora references