Mapika / decider-31b

decider-31b: typed decisions with calibrated probabilities on Gemma-4-31B, one pass, under one second

decider-31b answers typed decisions with a probability for every option. It reads each question in one forward pass and generates no text.

• Input: a state and one or more typed questions (Choice, Noul yes/no, Score), each with an explicit option list, in the System One wire format ( POST /v1/systemone ).
• Output: a probability distribution over the options of every question. Nothing needs parsing.
• Weights: google/gemma-4-31B-it with a fine-tune merged into the attention and MLP projections, quantized to NVFP4 (MLP weights and activations in NVFP4, attention projections in bf16, FP8 KV cache). About 31 GB.
• Code: decider.serve_vllm in decider-ai 1.9.0. It reads this repository's decider_config.json , which holds the configuration measured below.
• No thinking. This model never generates a reasoning trace. Every request is answered in one or two forward passes.

Results

Decision Index 0.3, public suite, all 140,178 requests (kit --edition 0.3 ), run by us on 2026-10-07/08 with exactly this repository (commit cf50c0b) and decider-ai 1.9.0 from PyPI, one server per B300. The board's full score also includes private tests that only the maintainers run; this is the public part only.

Public-index scores of other entries, as shown on the board on 2026-10-06 (their own runs, not ours):

Notes on the numbers:

• The run was not verified by the Decision Index maintainers and is not on the board.
• ECE_bw is the benchmark-weighted expected calibration error over the 214,304 scored fields whose rows are in the suite's selected-rows file. Rows outside that file are not part of the calibration figure.
• Results: Mapika/decision-index-results, runs/decider-31b (gated). Our development server gave 64.23 and ECE_bw 0.035 on the same suite; the figures above are from the released code.
• The temperatures were fitted on our own rows only. No Decision Index row was used to fit or choose them.
• Training used the public train splits of several benchmarks that the index also tests (see Training data). No Decision Index test row was used: every training row was compared with all suite rows and dropped on a match.

Latency

Measured one request at a time over HTTP on one B300, 400 requests drawn from the Decision Index 0.3 suite:

The Decision Index measures latency on an RTX PRO 6000. We have not measured on that GPU. Our previous entry (stock Gemma-4-31B) measured 3.2 times slower there on the mean than on our B300; on that factor this model's mean and p80 stay below one second. Treat that as an estimate.

Usage

vLLM 0.29.0 pins its own torch, so install in a separate environment:

python -m venv decider-vllm && . decider-vllm/bin/activate
pip install vllm==0.29.0 fastapi "uvicorn[standard]" jinja2 huggingface_hub
pip install --no-deps "decider-ai>=1.9.0"
DECIDER_MODEL=Mapika/decider-31b uvicorn decider.serve_vllm:app --host 127.0.0.1 --port 8000

It needs one GPU with NVFP4 support (Blackwell: B200/B300, RTX PRO 6000, RTX 50 series with enough memory).

import httpx
r = httpx.post("http://127.0.0.1:8000/v1/systemone", timeout=60, json={
    "state": {"ticket": "Order #1182 arrived broken. Customer asks for a refund; order is 40 days old; policy: 30 days."},
    "questions": {"refund": {"type": "noul", "instructions": "Is the refund allowed under the policy?",
                             "criteria": {"true": "allowed", "false": "not allowed"}}}})
print(r.json()["answers"]["refund"])

How a request is answered

1. First reading. Every question is rendered in the model's chat template with thinking off, and the log-probabilities of the option letters are read at the answer slot. A softmax at T1(n) = 2.042 + 0.003 ln n (n = number of options) gives the probabilities.
2. Second reading, only when unsure. A question whose top probability is below 0.7 is read again with its options in reversed order. The two log-softmaxes are averaged and a softmax at T2(n) = 1.953 + 0.010 ln n gives the final probabilities. On the Decision Index public suite this touches 28% of the questions; it removes part of the option-order bias at a small latency cost.
3. Questions above 0.7 keep the first reading.

The thresholds and temperatures are in decider_config.json ( temperature_by_options , second_reading ). They belong to these weights. DECIDER_SECOND_READING_BELOW=0 turns the second reading off (one reading per question; slightly faster, less accurate). The config also selects vLLM's Triton attention backend ( vllm_attention ): with this checkpoint's FP8 KV cache, the FlashInfer backend returned NaN in batched readouts in our tests.

Training data

The fine-tune was trained on decision items only, in the same layout as serving. The data:

• New for this model: rows from public train splits of tasks where our previous entry was weak, rendered in the Decision Index request format, and decontaminated against every Decision Index 0.3 suite row:
  • CLINC150 (intent classification, including out-of-scope), train split;
  • iSarcasmEval (task A, English), train file;
  • VAST (stance), train split;
  • ACOS (aspect-category sentiment, laptop and restaurant), train splits;
  • POP909-CL (chord labels), songs not used by the index;
  • GSM8K, train split; NLI4CT, train split;
  • generated smart-home requests (device states, household policies) from our own generator, with homes, rooms, devices and request texts drawn at random.
• From our earlier decider models: generated state-tracking decisions and decision families written by our own code, document questions written by Qwen3.6-27B, and labelled rows from public datasets (MMLU, ARC, CommonsenseQA, BoolQ, MNLI, SNLI, BANKING77, RACE, OpenBookQA, LogiQA 2.0, MedQA, WinoGrande).

Not used: Decision Index test rows, JevBench items.

Limitations

• One pass only. Tasks that need several steps of reasoning (GPQA 0.35, ChessBench 0.16, HLE 0.00 skill on the index) are not solved by reading harder. A thinking variant is a separate model.
• Calibration: ECE_bw 0.039 on the index. The model is underconfident on the tasks it was trained on (language area: accuracy 0.85 at confidence 0.74) and closer to calibrated elsewhere. Check calibration on your own data before relying on the probabilities.
• Trained skills vs. new tasks. On a 4,500-request index sample, compared with our previous readout adapter, the 8 benchmarks whose train splits were used gained 17 points of skill on average; the other 29 were unchanged on average (+0.6), with losses of 0.03 to 0.09 on SATA-Bench, Amazon ESCI, HoVer, GPQA and ChessBench (about 88 requests each, so partly noise).
• Hardware: NVFP4 needs a Blackwell GPU. We measured only on B300.
• vLLM version: the server patches a vLLM 0.29 internal and refuses other versions.
• Independent questions only: /decide and the schema cache of decider.serve are not served.

License

The weights are released under Apache 2.0, as the base model. Training sources and their terms:

• CLINC150: CC BY 3.0 (Larson et al., 2019). Credit: "An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction", clinc/oos-eval.
• iSarcasmEval: MIT (Abu Farha et al., 2022). The texts are tweets.
• POP909-CL: MIT; built on POP909 (Wang et al., 2020), MIT.
• VAST (Allaway and McKeown, 2020) and ACOS (Cai et al., 2021): no licence stated by the authors.
• GSM8K (Cobbe et al., 2021): MIT.
• NLI4CT (SemEval-2024 Task 2, Jullien et al.): no licence file in the task repository; released for the shared task.
• RACE: non-commercial research use. It is a small part of the replayed data (250 rows).
• ARC, SNLI, LogiQA 2.0 (copy dependent): CC BY-SA 4.0; BoolQ: CC BY-SA 3.0; BANKING77, MedQA: CC BY 4.0; MMLU, CommonsenseQA: MIT; MNLI: mixed per genre.