64 2 days 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-v3:q8_0

Details

2 days ago

4e20db734e8a · 13GB

gemma4
·
11.9B
·
Q8_0
clip
·
52.4M
·
F16
{ "num_ctx": 8192, "stop": [ "<turn|>" ] }

Readme

Decisio v3 (12B)

A Gemma 4 12B tuned for decisio, the open-source server that takes a text and typed questions (yes/no, one choice among options, a score on a scale you define) and returns a probability for every option from one forward pass per question.

What it is good at

It answers closed questions about a text noticeably more accurately than the base model and than the previous tuned 12B. Its probabilities are calibrated: when it says 80%, it is right about 80% of the time, at the temperatures built into these files. On 5,000 held-out questions from the distribution it was trained on it is right 78.8% of the time (base Gemma 4 12B: 67.7%, previous tuned 12B: 71.6%).

Tags

tag file size use it when
latest, q8_0 Q8_0 12.7 GB you want the closest match to the full-precision model
q4_k_m Q4_K_M 7.4 GB memory is tight and a small loss is fine

Both tags carry decisio’s decision keys and the model’s own temperatures, so the probabilities come out calibrated without extra settings.

How closely the quantised files follow the full-precision weights

Measured on 1,000 held-out questions, same text and same readout, against the merged bf16 weights:

same top answer median change in probability accuracy (bf16: 76.6%)
Q8_0 97.2% 0.0047 76.9%
Q4_K_M 93.6% 0.0176 76.2% (difference -0.4, 90% lower bound -1.5)

The Q8_0 file is within a hair of full precision. The Q4_K_M file changes the top answer on about one question in sixteen, and its accuracy stays within two points of bf16.

Use it

bash ollama pull aminroudaki/decisio-v3 # the 8-bit file ollama pull aminroudaki/decisio-v3:q4_k_m # the 4-bit file `

Then point decisio at the Ollama endpoint; the /v1/systemone route answers with a probability for every option.
Context is set to 8,192 tokens in the Modelfile.

Limits

  • The figures are readings on held-out questions of the training distribution, not scores on an independent benchmark.
  • Other servers that apply no temperature give flatter or sharper probabilities than the calibrated ones stated here.
  • Through Ollama the model answers on its own: Ollama’s decision route does not run decisio’s server, so the calculator block is not added.
    The calculator (exact values for bills, invoices, payslips, dates and times) comes with decisio’s own server (0.51.1 or later), on a GPU with vLLM or on a Mac with the MLX repositories; decisio fetches the small tagger it needs by itself.
    If your texts carry numbers to add up, use that route: it is noticeably more accurate on them.
  • Images are not part of this model’s training.

Apache License 2.0, as the base model’s.
Built by Tachara AI Lab.