jrlabs01/ s1:latest

8 2 days ago

s1: an open, calibrated System One decision model

vision tools thinking 
ollama run jrlabs01/s1

Details

2 days ago

3fd58603b806 · 18GB

json
{ "architectures": [ "Gemma4ForConditionalGeneration" ], "audio_config": null, "audio_token_id": 258
json
{ "bos_token_id": 2, "do_sample": true, "eos_token_id": [ 1, 106, 50 ], "pad_token_id": 0, "temperat
json
{ "audio_ms_per_token": 40, "audio_seq_length": 750, "feature_extractor": { "dither": 0.0, "feature_
json
{ "version": "1.0", "truncation": null, "padding": null, "added_tokens": [ { "id": 0, "content": "<p
json
{ "audio_token": "<|audio|>", "backend": "tokenizers", "boa_token": "<|audio>", "boi_token": "<|imag
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "temperature": 0 }
1013 tensors

Readme

s1: an open, calibrated System One decision model

s1 makes typed decisions over text or JSON (pick an option, answer yes/no, or place something on an ordered scale) and gives a calibrated probability for every option, from one forward pass. This is the int4 build of j-raghavan/s1-gemma4-26b-decision, a fine-tune of Gemma 4 26B-A4B. Apache-2.0.

Accuracy of this build (held-out test sets): JevBench subset 0.801 (bf16: 0.808), structured decisions 0.966 (bf16: 0.971); neither difference is significant. About 26 GB of memory while loaded.

Use it the right way

s1 reads the probabilities of the option letters after the prefix Answer:. Send raw prompts that start with <bos>; Ollama’s chat formatting changes the prompt and the answers.

The easiest way is the repository’s HTTP API, which does this for you and adds calibration:

ollama pull jrlabs01/s1
git clone https://github.com/j-raghavan/s1-decision-model && cd s1-decision-model
S1_MODEL=jrlabs01/s1 uv run --extra api uvicorn api.server:app --port 8000
curl -s localhost:8000/v1/decisions -H 'content-type: application/json' -d '{
  "state": {"ticket": "I was charged twice for order 4471."},
  "questions": {"team": {"type": "choice", "instructions": "Which team should handle this?",
    "criteria": {"billing": "Payments and refunds", "shipping": "Deliveries", "tech": "Bugs"}}}}'

Calling Ollama directly: POST /api/generate with "raw": true, "logprobs": true, "top_logprobs": 20, "options": {"num_predict": 1, "temperature": 0} and the prompt format in examples/quickstart.py, prefixed with <bos>.

Benchmarks, methodology and training data: github.com/j-raghavan/s1-decision-model. s1 is an independent project, not affiliated with TypeSafe AI (Jev).