Decisio-12b-tuned-mlx8-commit

The 8-bit MLX version, for Apple silicon, of tachara-ai/Decisio-12b-tuned-commit: the Gemma 4 12B tuned as a model that answers typed questions about a state with a probability for every option, the commit flavor. It is one of two repositories of the same weights (Decisio-12b-tuned-mlx8-calibrated and Decisio-12b-tuned-mlx8-commit); they differ in decision_config.json only.

Serving

uv sync --extra mlx    # or: pip install "decisio[mlx]"
uv run python -m decisio.serve.vllm_engine --backend mlx --base tachara-ai/Decisio-12b-tuned-mlx8-commit

Needs decisio 0.34.0 or later. A plain request gets the repository's defaults with no flag and no adapter field (/health, systemone.repository_defaults and systemone.flavor): the header layout, the temperatures per question type and the flavor's settings. The server is named Decisio-12b-tuned-mlx8-commit. The weights are 12.7 GB. mlx-lm alone loads them but does not apply decisio's prompt layout, option-letter readout, temperatures or commit rule.

What this repository is

The calibrated probabilities, with the yes/no commit rule on top: a yes/no probability strictly between 0.20 and 0.80 is reported at the edge of that band on its own side (0.80 above one half, 0.20 at or below it). The answer never changes, its calibration does.

Weights

MLX affine quantization, 8 bits per weight in groups of 64, made by mlx-lm 0.32.2 from the bf16 weights of [tachara-ai/Decisio-12b-tuned-commit]: the adapter was merged into the base in fp32 and rounded once to bf16 (stochastically, seed 0) before this conversion. The conversion is not decisio's own; decision_config.json lists every file by sha256 and says so, and SHA256SUMS lists the files of this repository. Measured on 1,894 questions the model was not trained on, against the bf16 repository served by decisio with the same settings: the top answer is the same on 97.9% of them (yes/no 99.7%, choice 97.4%, score 98.1%), and accuracy is 69.8% against 69.9% (a difference of -0.1 points; paired 90% interval -0.5 to +0.3). On a 100-question check the largest difference in a probability from bf16 was 0.048 (mean 0.0095) and all 100 top answers were the same. A fresh server start answers a 150-question sample identically. Memory, measured in decisio's MLX engine on an Apple M5 Pro with 64 GB, with no swap used: the weights take 13.8 GB and the peak is 14.9 GB at 8k tokens of state, 16.2 GB at 16k and 18.6 GB at 32k (16.4 GB at 32k with ten questions); a 32 GB Mac has room. A 6-bit version, Decisio-12b-tuned-mlx-calibrated and Decisio-12b-tuned-mlx-commit, takes less memory and is a little further from bf16. The temperatures were fitted by minimum log loss per question type and choice band on the bf16 model's plain readout and are used unchanged.

Data

Training data not disclosed. No Decision Index or JevBench evaluation items were used in training or checkpoint selection; the data was decontaminated against both boards' public test sets.

Training included the official training splits of some public benchmarks, used under their licences. No test or evaluation items of those benchmarks were used, and every training row was checked against both boards' public test sets.

Licence

The licence of the base model, Apache License 2.0 (see LICENSE and NOTICE). Gemma 4 is Google's.