Decisio-12b-tuned-commit-think

The Gemma 4 12B tuned as a model that answers typed questions about a state with a probability for every option: the commit-think flavor. This is google/gemma-4-12B-it (revision 707f0a3b) with a LoRA adapter merged into its weights, served with decisio. It is one of three repositories of the same weights (Decisio-12b-tuned-calibrated, Decisio-12b-tuned-commit and Decisio-12b-tuned-commit-think); they differ in decision_config.json only.

Serving

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

Needs decisio 0.35.0 or later; an older decisio does not know the repository's flavor. A plain request gets the repository's defaults with no flag and no adapter field (shown in /health, systemone.repository_defaults and systemone.flavor): the header layout, the temperatures per question type (yes/no 3.314, score 2.1109, choice 1.8222; choice by number of options 2-10: 2.1317, 11-26: 1.8766, 27-255: 1.7181), the yes/no commit rule. The server is named Decisio-12b-tuned-commit-think; every flag that differs from the repository is reported as an override. vLLM alone loads the weights as an ordinary Gemma 4 checkpoint but does not apply decisio's prompt layout, option-letter readout, temperatures or commit rule.

What this repository is

The commit flavor (the calibrated probabilities, with the yes/no commit rule on top) that also thinks on a few choice questions: a choice question whose served top probability is under 0.3712 is read again after the model has written up to 256 tokens of reasoning, and the two reads are mixed; no other question type thinks. A state longer than 45,000 tokens is cut to its start and answered from there, without thinking, and the response says it was cut.

Weights

bf16. The adapter was merged into the base's weights in fp32 and rounded once to bf16, stochastically with seed 0, by python -m decisio.hub_bf16 merge; the merge keeps 0.999996 of the adapter's change in weight space, and every other tensor is the base's, byte for byte. decision_config.json lists every file by sha256 and the merge facts; SHA256SUMS lists the files of this repository. The temperatures were fitted by minimum log loss per question type and choice band, on the plain readout. Measured on questions screened against our training data, against the same model served without thinking. Choice questions whose top probability is under 0.3712 are read again after up to 256 tokens of reasoning; no other question type thinks. On hard multi-domain decisions (525 questions) the top answer is right on 4.95 points more (90% interval 3.43 to 6.67; 27 fixed, 1 broken) and 8.95% of the questions think; the settings were chosen on these questions, so that gain is optimistic. On natural-wording questions (939), which were not used to choose them, it is 0.75 points more (0.21 to 1.38; 9 fixed, 2 broken) and 2.56% think. On a mixed set of development items (430) no question thinks. Across 2,044 decisions none took more than 10 seconds (the slowest 4.71 s on one RTX PRO 6000) and the median time of a decision is unchanged (60 ms); a thought is a median of 170 tokens, 29% of them reach the 256-token cap. Choice calibration is unchanged within noise (error 0.048 single pass, 0.050 with thinking). A state longer than 45,000 tokens is cut to its start, the response says so, and the question is answered from there by the single pass, without thinking. On 60 long documents with the deciding facts at the start, middle or end, 39 were right within 10 seconds with the cut (40 were right in all: 17 of 20 at the start, 14 in the middle, 9 at the end; one right answer took longer than 10 seconds), and a cut state's first read took about 7.5 s. The other settings of this repository (the header layout, the temperatures, the commit rule) are those of Decisio-12b-tuned-commit, whose weights this repository shares.

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.