828 Downloads Updated 1 week ago
curl http://localhost:11434/v1/systemone \
-H "Content-Type: application/json" \
-d '{
"model": "aminroudaki/decisio-gemma:q8_0-ollama0.40.3",
"state": "Hello World",
"questions": {
"says_hello": {
"type": "noul",
"instructions": "Does the state text contain a greeting?",
"criteria": {
"true": "The state text contains a greeting.",
"false": "The state text does not contain a greeting."
}
}
}
}'
decisio-gemma answers closed questions about any text or image with Gemma 4 12B, with a 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 uses the same prompt and temperature as the decisio server, and on the questions measured its q8_0 tag chooses the same answer as decisio’s own server for 98% of them.
Requires Ollama 0.40.4 or later (checked on 0.40.4, 0.40.9 and 0.40.16).
ollama pull aminroudaki/decisio-gemma
curl http://localhost:11434/v1/systemone -d '{
"model": "aminroudaki/decisio-gemma",
"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:
curl http://localhost:11434/v1/systemone -d '{
"model": "aminroudaki/decisio-gemma",
"state": "The customer sent this photo of the parcel.",
"images": ["'"$(base64 < parcel.jpg | tr -d '\n')"'"],
"questions": {"damaged": {"type": "noul", "instructions": "Is the parcel damaged?"}}
}'
| Tag | Download | Memory while answering (8,192-token context) |
|---|---|---|
q4_k_m (also latest) |
7.8 GB | 11.1 GB |
q8_0 |
12.8 GB | 15.9 GB |
q4_k_m-ollama0.40.3, q8_0-ollama0.40.3 |
the earlier files, for Ollama 0.35.1 through 0.40.3 |
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.
systemone chat template. For a request of one question with at most 68 options it is the decisio server’s prompt byte for byte.On 300 questions from decisio’s test suite, the q8_0 tag chose the same answer as decisio’s own server on the same Mac for 98.3% of them (q4_k_m: 87.0% against decisio’s 4-bit conversion, about as often as decisio’s own 4-bit and 8-bit conversions agree with each other), and on 257 image questions both tags chose decisio’s answer for 97% to 98%.
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; - decisio’s commit flavor, tasks, abstention and the fast table are not available here.
For those, run decisio itself with --base gemma-4-12b: github.com/aminry/decisio.
The tuned 12B is aminroudaki/decisio-gemma-tuned; for the Qwen base, see aminroudaki/decisio (Ollama 0.35.1 through 0.40.3).