Jet-0.6B

Jet is a family of decision models. This repository contains Jet-0.6B (V5), the last Qwen3-0.6B release. The current model is Jet-4B. This release was previously published in the history of quaedra/jet (now quaedra/jet-4b) at revision 25ccbd9e09c75643b3c2214e2b2522bec39171a7; only this model card's links changed.

Jet is a small decision model built on Qwen3-0.6B. Give it a state and named, typed questions; it returns choices, scores, and probabilities without generating free-form text. Answers always follow the requested type, but decisions can still be wrong.

Model weights · Training history

Type	Criteria	Answer
`choice`	2–255 named options	Selected key, probability per key, confidence
`score`	2–10 ordered levels	Fractional score, selected level, probabilities
`noul`	None	Probability that the answer is yes

Run locally

git clone https://github.com/quaedra/jet
cd jet
uv sync                  # Apple Silicon / Metal
# Linux with NVIDIA: uv sync --extra cuda
JET_API_KEY=secret uv run jet-serve --base-model quaedra/jet-0.6b

curl http://localhost:8000/v1/decide \
  -H 'Authorization: Bearer secret' \
  -H 'Content-Type: application/json' \
  -d '{"state":"I was charged twice this month.","questions":{"topic":{"type":"choice","instructions":"What is the primary issue?","criteria":{"billing":"billing or payment problem","bug":"the product is broken"}},"escalate":{"type":"noul","instructions":"Does this require human support?"}}}'

The model requires Jet's prompt format and label-token readout. It is not a chat model. The MLX server shares the state prefix across questions, applies the saved calibration temperatures, and returns typed answers. Serving truncates long states in the middle; the Decision Index adapter instead requires complete inputs.

How it works

Each answer option maps to a single token. Jet reads the next-token logits only for those labels and applies softmax with a temperature fitted on held-out data. Standard serving uses one forward pass per question, without sampling. The fused bf16 weights include the trained LoRA updates.

The benchmark package also provides decision_index_ensemble:TwoOrderJetEngine. It averages original and reversed option-order probabilities using exact option keys. This approximately doubles forward-pass work and is separate from the standard serving API.

Training and evaluation

The current checkpoint was trained locally on an RTX 4080 Super with MLX CUDA. The final training stage used 15,997 examples, rank-16 LoRA, learning rate 1e-5, gradient checkpointing, and one epoch. Selection chose step 3,500 of 3,570. Public training partitions and deterministic generators cover relevance, entailment, stance, sarcasm, routing, arithmetic, Boolean rules, code behavior, commonsense completion, and tool-response preference. Calibration uses a separate held-out split.

Held-out set	Rows	Accuracy
General typed decisions	3,683	66.66%
Intent, entailment, commonsense and tool-response preference	600	73.50%
Transfer tasks	720	67.50%
Domain-held-out tool routing	240	85.00%

These are single-order adapter measurements. They are not a Decision Index leaderboard score. Partial public benchmark reconstructions do not cover the complete current suite. Cross-dataset text overlap inherited during training also prevents a certified decontamination claim. Performance varies by task; see the evaluation report for full methodology, limitations, and the separate two-order results.

To reproduce training, source acquisition, audits, and evaluation, use the recorded protocol and scripts under scripts/. Dataset sources, revisions, exclusion checks, configuration, and checkpoint hashes are recorded under docs/training/.

# CUDA launcher prepares the MLX CUDA environment.
./train_cuda.sh --help
uv run jet-calibrate --adapter /path/to/adapter --data /path/to/calibration.jsonl
uv run jet-fuse --adapter /path/to/adapter --out models/jet
uv run jet-eval --base-model models/jet --data data/test.jsonl

Optional jet-distill data generation uses Anthropic credentials or the --backend claude-code mode. It is not required for serving or reproducing the public-source training pipeline; hosted generation can incur charges.

Deployment

The Hugging Face model package includes fused bf16 weights, tokenizer, calibration, golden reference vectors, and a q8 ONNX export. The Gradio Space uses a pinned model revision and exposes /decide; its availability depends on Hugging Face's free hosting quota. See deployment instructions.

The ONNX model supports CPU and browser runtimes through ONNX Runtime. jet-golden generates reference cases, and export_web.sh exports and checks the browser model. Export validation is recorded with the model package; quantization can change probabilities.

Project layout

src/format.py, src/inference.py: prompts and typed answers
src/model.py, src/torch_model.py, src/onnx_model.py: inference backends
src/train.py, src/evaluate.py, src/fuse.py: training, calibration, evaluation and fusion
src/data/, scripts/: data builders and reproducible experiments
src/decision_index_engine.py, src/decision_index_ensemble.py: benchmark adapters
deploy/huggingface/: hosted demo and API deployment