any2jev-qwen3-0.6b

A Jev-style System One decision model made with any2jev from Qwen/Qwen3-0.6B. State in, typed Choice / Score / Noul answers with calibrated probabilities out, in one forward pass. No text is generated.

pip install "any2jev[serve]"
any2jev serve hf://huaweifeng/any2jev-qwen3-0.6b      # POST /v1/systemone, TypeSafe SDK compatible
any2jev ask hf://huaweifeng/any2jev-qwen3-0.6b --state "My payouts have failed 3 days in a row, fix this ASAP" \
    --choice "Which team? | billing, technical, sales" --noul "Is this urgent?"

What is in this repo

• adapter/: LoRA adapter (r=16) on Qwen/Qwen3-0.6B, trained with the vocabulary head removed
• head.safetensors: the pointer head (dim 256) that scores options against the decision token
• any2jev.json: delimiters, mode (packed), temperature 1.609 fitted on validation
• tokenizer/: the base tokenizer (delimiter tokens reused or added)
• train_report.json, eval.json: training config, history and held-out metrics

Training

• data: data/public/train.jsonl; 10.6 M trainable parameters, 1.0 epoch(s), lr 0.0002, batch 4 x 2
• wall clock: 61 min on one consumer GPU

Held-out evaluation

group	n	accuracy	NLL	Brier	ECE	AURC
overall	1000	0.796	0.506	0.284	0.027	0.062
noul	250	0.844	0.386	0.242	0.090	0.058
choice	500	0.906	0.272	0.142	0.020	0.018
score	250	0.528	1.093	0.611	0.056	0.424

Option-order test: argmax stable in 96% of 23 Choice questions. Isolation check (packed vs separate questions): max |dp| = 0.0e+00.

Limitations

Trained on a few thousand labelled decisions from a handful of sources; expect the accuracy above on similar inputs and lower accuracy off-distribution. Probabilities are calibrated on the validation split, not a guarantee per answer. Keep arithmetic, dates and counting in code, as Jev's own docs recommend.

Independent project, not affiliated with TypeSafe AI. Apache-2.0.