Decisio-12b-tuned-fp8-commit

The FP8 version 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-fp8-calibrated and Decisio-12b-tuned-fp8-commit); they differ in decision_config.json only.

Serving

uv run python -m decisio.serve.vllm_engine --base tachara-ai/Decisio-12b-tuned-fp8-commit

Needs decisio 0.33.0 or later and a GPU with FP8 support. 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-fp8-commit. decisio 0.33.0 turns torch's precision-cast emulation on for FP8 bases (--emulate-precision-casts auto), which narrows the difference between two cold starts of the same FP8 model. vLLM alone loads the weights as an ordinary Gemma 4 FP8 checkpoint 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

FP8 (float8 e4m3) with per-channel weight scales and dynamic per-token activation quantization (FP8_DYNAMIC), made by llm-compressor 0.14.0.1 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 quantization. The quantization is not decisio's own merge; decision_config.json lists every file by sha256 and says so, and SHA256SUMS lists the files of this repository. Measured against the bf16 repository (Decisio-12b-tuned-calibrated) on questions screened against our training data, top answer against the key, paired difference in points with a 90% interval: hard multi-domain decisions (525 questions) 75.05 to 75.24 (+0.19, -1.14 to +1.52); mixed development items (430) 66.28 to 66.05 (-0.23, -1.63 to +1.16); natural wording (939) 69.01 to 69.22 (+0.21, -0.53 to +0.96). It is level with the bf16 repository on these sets (differences within 0.3 points, every interval includes zero), and it changes the top answer on about 3 to 6 of every 100 questions (the same top answer on 94.3%, 94.9% and 97.0% of the three sets), so it is not an answer-for-answer copy of bf16. The commit flavor gives the same top answer as the calibrated one on all 430 mixed development items, and its yes/no accuracy with the commit rule, 88.0% on 125 questions, equals the bf16 repository's. One pair of cold starts of the calibrated flavor answered 500 questions identically. 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.