SnapDragon-Coder

status

closed · mixed

when

2026-07 → 2026-08

came from

Greenfield DevOps platform, Dream Team, First public repos

led to

shellm, Marilyn, Fleet & operations

Turn 1: retrain GLM-4.5-Air into a better Go coder by distillation and verified-reward fine-tuning.

Why it started

The platform’s builder tiers, inherited from Dream Team, needed a local coding model that spoke the build dialect. A large open mixture-of-experts model was close. The question was whether a few hundred verified Go tasks and a reward signal could close the gap. The repo opened on 2026-07-11, five days after the platform cutover.¹

What it is

A distillation pipeline: a pool of 502 Go tasks with executable checks, 1,290 harvested traces, RLVR (reinforcement learning from verifiable rewards, where the sandbox grades the answer) and QLoRA fine-tuning against a quantised base, and a frozen 77-task benchmark to score each build. Served from the workstation card on monte. A second experiment, Domain-REAP, asked whether pruning the expert layers to a domain would keep quality at a fraction of the size.²

Where it stands

Closed. v1.0 reached 64 of 77 on my own frozen 77-task benchmark, matching the best open model I could run on that same set, and was a build-dialect specialist that collapsed outside its dialect. The decisive finding was that a community 4-bit quantisation, not the model family, was the quality killer: the benchmark’s op-adoption score 0.82 quantised versus 0.98 in bf16, and the adapter compounded it. Domain-REAP was closed with the full base beating every carve (63 vs 34 vs 23).³

Lineage

• Came from: greenfield platform, Dream Team’s builder tiers, and the 22-model benchmark in the first public repos
• Led to: shellm (if fine-tuning a big base handicaps it, train small from scratch), the GPU lease manager and sandboxing in fleet ops, and the RLVR trace corpus reused by Marilyn

Sub-pages

• Rabbit holes