system-one-qwen3-0.6b — volume-tier System One adapter

LoRA adapter for Qwen3-0.6B (base, Apache-2.0) turning it into a System One decision model: typed questions (choice / score / noul) over a JSON state, answered with calibrated probability distributions read directly from restricted logits in one forward pass — no text generation. Part of system-one, an open learning project rebuilding the behavior of Typesafe's Jev. Not affiliated with or endorsed by TypeSafe.

Files

• adapters.safetensors + adapter_config.json — rank-16 attention LoRA (MLX format)
• temperature.json — per-primitive temperatures (choice 1.019 / noul 1.156 / score 1.056), fitted on a dedicated calibration split and bound to these exact weights: the serving engine refuses them against anything else.

Training

Frozen recipe (decision 25 in the repo's audited decision log): 8.7k questions recast from public classification datasets plus ~1.6k synthetic scenarios whose probability targets were produced by pinned jev-1.13.0 (teacher distillation; disclosed deliberately) and filtered for teacher overconfidence. Readout-matched cross-entropy on target distributions + Ranked Probability Score on ordinal questions; LR 1e-5, batch 8, one epoch, seed 42, Apple MLX.

Evaluation (full protocol and CIs in the repo)

Split	Accuracy	ECE	Notes
Development (n=1,128)	82.0%	0.020	used for steering; optimistic
Reserved in-family test (n=1,048, spent once)	77.6%	0.049	unbiased; choice 80.6 / noul 84.7 / score 50.0
Fresh out-of-family test (n=1,048, spent once)	66.5%	0.076	new task families; see shift warning

Shift warning (measured): confidence thresholds are valid in-family only. On out-of-family workloads confident-error rates reached ~19% at t>=0.9 versus ~2% in-family. The serving runtime supports per-workload temperature fitting from ~100 labeled examples (POST /v1/calibrations), which repaired scalar miscalibration to 3-4% confident errors in the repo's resampled experiment.

Use

git clone https://github.com/mpuig/system-one && cd system-one && uv sync
hf download mpuig/system-one-qwen3-0.6b --local-dir system-one-qwen3-0.6b
uv run python scripts/serve.py --model Qwen/Qwen3-0.6B \
  --adapter system-one-qwen3-0.6b --temperature system-one-qwen3-0.6b/temperature.json

(The same adapter and temperature file also ship inside the GitHub repo at adapters/qwen3-0.6b-structured-v1-synthfiltered-rps and release/temperature-qwen3-0.6b.json, so the download step is optional there.)

Apple Silicon required (MLX). Score is this tier's weak primitive (50% on the unbiased test); use the quality tier for score-heavy workloads. This is a research prototype, not a certified production service.