wonder

A self-improvement training loop that fine-tunes a small local coding model (Qwen3-8B, 4-bit, on a single 12GB consumer GPU) to get measurably better at coding — by training only on its own verified-correct solution traces.

The loop has four parts. The student is Qwen3-8B running as a coding agent. The verifier is the test suite — automatic ground truth, and the only thing that promotes weights. The teacher is a stronger frontier model, invoked at checkpoints to demonstrate what the student cannot yet crack. The trainer is a separate QLoRA/SFT stage that consumes verified trajectories.

The objective is fitness = success ÷ tokens: solve more, with fewer tokens, on a private codebase's idioms.

Status

Past the first transfer gate: the loop teaches a transferable capability rather than memorising a task set. A taxonomy-withheld boundary that failed at v0 was healed by v0.2 and held on the held-out set — a capability stock Qwen3-14B, a 2.6×-larger model, still lacks. v0.3 was deliberately not promoted: it bought a real format-brittleness fix with real held-out erosion.

Why it is here

This site's entire design language comes out of wonder's design/ lane — the HUD v4 design brief, the honest-data law, the 8×16 character grid, the tmux-style pane titles. The dossier is the skin; wonder wrote the rules underneath it.

Repository