43 18 hours ago

Decision Model Tuned: Typed questions about text or images, Answers yes/no, multiple-choice and scoring questions about any text, returning a probability for every option.

vision tools thinking
ollama run aminroudaki/decisio-gemma-tuned:q4_k_m

Details

18 hours ago

6c30098bd5e8 · 7.5GB

gemma4
·
11.9B
·
Q4_K_M
clip
·
52.4M
·
F16
{ "num_ctx": 8192, "stop": [ "<turn|>" ] }

Readme

decisio-gemma-tuned

decisio-gemma-tuned answers closed questions about any text or image with a tuned Gemma 4 12B, with a calibrated probability for every option. Give it a piece of text, optionally images, and one or more questions (yes/no, a choice among up to 255 options, or a score on a scale you define); it returns, for each question, the chosen answer and how likely each option is, from one forward pass per question. It is Decisio-12b-tuned-calibrated with the decisio server’s prompt and fitted temperatures, and on the questions measured its q8_0 tag chooses the same answer as decisio’s own server for 99% of them.

How to use it

Requires Ollama 0.40.4 or later (checked on 0.40.4, 0.40.9 and 0.40.16).

ollama pull aminroudaki/decisio-gemma-tuned
curl http://localhost:11434/v1/systemone -d '{
  "model": "aminroudaki/decisio-gemma-tuned",
  "state": "Order 4471 shipped on Monday. The customer writes: My parcel arrived with the box crushed and the lamp inside broken.",
  "questions": {
    "intent": {"type": "choice", "instructions": "What does the customer want?",
               "criteria": {"track_order": null, "report_damage": null, "cancel_order": null, "change_address": null}},
    "refund": {"type": "noul", "instructions": "Is the customer likely to ask for a refund or replacement?"}
  }
}'

Images go in the request’s images field, base64-encoded.

Tags

Tag Download Memory while answering (8,192-token context)
q4_k_m (also latest) 7.5 GB 10.8 GB
q8_0 12.8 GB 15.9 GB

Memory was measured on a Mac as the resident set of Ollama’s runner: the weights, mapped from the file, and about 3.5 GB of buffers.

What is in it

  • Weights: tachara-ai/Decisio-12b-tuned-calibrated at revision baa731aa (Apache-2.0): google/gemma-4-12B-it with a LoRA adapter merged into its weights, converted with llama.cpp and quantised with an importance matrix, with Gemma 4’s vision projector (unchanged by tuning).
  • Prompt: decisio’s own prompt for each question, carried in the file as its systemone chat template. For a request of one question with at most 68 options it is the decisio server’s prompt byte for byte.
  • Calibration: the model’s fitted temperatures, the ones the decisio server applies to it: yes/no 3.314, score 2.1109, choice 2.1317 for up to 10 options and 1.8766 from 11 (decisio uses 1.7181 from 27 options; this file uses 1.8766 for every count from 11).

How it compares with the decisio server

On 300 questions from decisio’s test suite, the q8_0 tag chose the same answer as decisio’s own server for 99.0% of them and the q4_k_m tag for 94.7%; on 257 image questions both chose decisio’s answer for 97%. No accuracy difference was outside its 95% interval.

Where it differs from the decisio server: - in a request of several questions, each question’s prompt lists only its own question; - images go before the text, apply to the whole request, and are read near their own resolution; there is no Box question; - the commit flavor (Decisio-12b-tuned-commit), tasks, abstention and the fast table are not available here.

For those, run decisio itself with --base tachara-ai/Decisio-12b-tuned-calibrated: github.com/aminry/decisio. The Gemma 4 12B base is aminroudaki/decisio-gemma.