chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF

Jev-Style v3 is available — smaller and stronger: Jev-Style-0.8B-Decision-v3-GGUF scores 79.2% on the 2,000 typed decisions (v1: 53.4%, v2: 73.5%, as reported on the v2 card), takes 25,600-token inputs, works across 51 languages and scores options without the 26-letter cap (tested with 77 options), all at 0.8B parameters. This repository preserves v1; v2 is here.

Jev-Style-Qwen3.5-2B-Decision (GGUF)

Website: jevstyle.com — all JevStyle decision models, benchmarks and quickstart in one place.

A Jev-style decision model: it does not write text. Give it a state, a question and a list of options, and it returns the decision with calibrated probabilities from a single token position. Runs in LM Studio and llama.cpp.

File	Size	Same decision as bf16	Accuracy (500 held-out)
Jev-Style-Qwen3.5-2B-Decision-Q4_K_M.gguf	1.3 GB	94.4%	82.4%
Jev-Style-Qwen3.5-2B-Decision-Q8_0.gguf	2.1 GB	99.4%	81.6%
Jev-Style-Qwen3.5-2B-Decision-BF16.gguf	3.9 GB	99.8%	81.6%

MLX bf16 build for Apple Silicon: chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-MLX-bf16

Results

Everything below is measured on data the model never trained on, with the probabilities exactly as the released weights produce them (no post-processing).

	Qwen3.5-2B-Base, zero-shot	This model
Accuracy, 5 decision tasks (1,500 held-out examples)	65.9%	82.3%
Calibration error (ECE) on those tasks	0.065	0.017
Negative log-likelihood / Brier score	0.786 / 0.446	0.418 / 0.242
Calibration error on task types never seen in training	0.155	0.075
Latency per decision (M1 Max, MLX bf16)	76 ms	77 ms (no added cost)

• Calibrated out of the box. An ECE of 0.017 on 1,500 examples is statistically indistinguishable from a perfectly calibrated model: simulating labels from the model's own probabilities gives an expected ECE of 0.017 (95th percentile 0.025) from sampling noise alone. When this model says 80%, it is right about 80% of the time.
• Large accuracy gains where the base model struggled: MNLI 52.3% -> 86.7%, SST-5 32.0% -> 61.7%, BoolQ 73.0% -> 82.7%, SST-2 87.3% -> 92.7%, AG News 84.7% -> 87.7%.
• Calibration transfers to new task types: on emotion classification and RTE (never seen in training) the calibration error is halved (0.155 -> 0.075) at unchanged accuracy (64.5%).
• Zero-cost calibration. A temperature fitted on 4,366 held-out examples is folded into the final RMSNorm weight, so every logit is already calibrated. Nothing to apply at inference time.
• Quantisation-friendly. Q8_0 makes the same decision as bf16 on 99.4% of examples; Q4_K_M (1.3 GB) keeps 82.4% accuracy.
• Efficient training recipe. LoRA rank 16 on all linear layers with a log-score loss, built on a custom chunk-parallel, differentiable Gated DeltaNet forward that matches the per-token training path to 1e-6 (outputs, state and all gradients) and is 6.5x faster per step (measured on the 0.8B sibling model).

What "Jev-style" means

Jev (TypeSafe AI, 2026) introduced System One models: instead of generating text, the model takes a state plus a typed question and returns a decision with calibrated probabilities in a single pass. This model follows that pattern on top of an open base model:

• Choice - pick one of N declared options, with a probability for each
• Bool - probability that a proposition is true
• Score - a distribution over ordered levels, and its expectation as a continuous score

It cannot answer outside the declared options, it does not decode text, and one prefill pass gives the whole distribution.

This is an independent, from-scratch reproduction of the publicly described idea. It is not affiliated with TypeSafe AI and is not the Jev model.

Quick start

Important: you must use the prompt format below. Plain chat messages produce meaningless text continuation. This is a decision function, not a chat model. In a chat window, paste the full prompt (it must end with Answer:) and start a new chat for every decision.

LM Studio - download a file from this repo, load it, start the local server (Developer tab), then:

python jev_style_client.py --url http://localhost:1234 --model jev-style-qwen3.5-2b-decision

llama.cpp

llama-server -hf chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF:Q8_0 --port 8080
python jev_style_client.py --url http://localhost:8080

jev_style_client.py (standard library only) asks for one token with top_logprobs on /v1/chat/completions and renormalises the option letters:

from jev_style_client import decide, decide_bool, decide_score

decide("http://localhost:1234",
       "Shares of the chipmaker jumped 8% after it raised its revenue forecast.",
       "Which news section does this article belong to?",
       ["World", "Sports", "Business", "Science/Technology"],
       model="jev-style-qwen3.5-2b-decision")
# [('Business', 0.68), ('Science/Technology', 0.31), ('World', 0.005), ('Sports', 0.002)]

Verified end to end through LM Studio's server on 500 held-out examples: 81.6% accuracy, ECE 0.028, about 110 ms per decision over HTTP on an M1 Max. In the LM Studio chat window you can also paste the prompt below and the model replies with the option letter.

Prompt format

You are a decision function. Read the state, then answer the question by choosing exactly one option.

[State]
{state}

[Question]
{question}

[Options]
A. {option 1}
B. {option 2}

Answer:

The next token is the option letter ( A, B, ...). Its probability, renormalised over the declared letters, is the decision distribution. For Score, list the levels in order; for Bool, use yes / no. The repository ships a pass-through chat template, so chat endpoints and the LM Studio chat window pass this text to the model verbatim.

Scope

• A decision function, not a chat model: send the prompt format above.
• Trained on five English task families (sentiment, natural-language inference, topic, yes/no question answering, 5-level rating). On unseen task types it keeps the base model's accuracy with better, though not perfect, calibration.
• Up to 26 options (20 when probabilities are read through a server's top_logprobs).

Training data and licence

SST-2 and MNLI (GLUE), AG News, BoolQ and SST-5, 22k examples converted to typed decisions; 80% for LoRA training, 20% held out for the calibration temperature. AG News is distributed for research / non-commercial use. Weights: Apache-2.0, same as Qwen/Qwen3.5-2B-Base.

Contact

I welcome internship, employment, and research collaboration opportunities. Please contact me at [email protected].

欢迎提供实习、工作及科研合作机会，请邮件联系：[email protected]。