j-raghavan/s1-gemma4-26b-decision
j-raghavan/s1-gemma4-26b-decision is an open, calibrated System One (s1) decision model based on google/gemma-4-26B-A4B-it, a 26 billion parameter mixture-of-experts architecture with approximately 4 billion active parameters. Fine-tuned with LoRA on 100,000 typed decision rows, this model provides probabilities for every option in a single forward pass given a state and a typed question, without generating text. It excels at structured decision-making tasks, offering calibrated probability outputs for choice, yes/no, and scoring questions.
Loading preview...
s1-gemma4-26b-decision: A Calibrated Decision Model
s1-gemma4-26b-decision is an open, calibrated System One (s1) decision model, fine-tuned from Google's gemma-4-26B-A4B-it (a 26B parameter Mixture-of-Experts model with ~4B active parameters). Unlike traditional LLMs that generate text, s1 provides a probability for every option in a single forward pass, making it highly efficient for structured decision tasks.
Key Capabilities
- Direct Probability Outputs: Given a state (text or JSON) and a typed question, it returns probabilities for each option without generating conversational text.
- Multiple Question Types: Supports
choice(keyed options with descriptions),noul(yes/no conditions), andscore(ordered levels with probabilities and expected level). - High Performance: Achieves strong results on held-out test sets, including 0.817 (calibrated) on JevBench subsets and 0.981 (calibrated) on custom structured decisions.
- Efficiency: Offers low latency, with median decision times of 28 ms (FP8) or 40 ms (bf16) on an NVIDIA RTX PRO 6000.
- Quantized Support: Available in int4 quantization for local deployment (e.g., via Ollama), enabling use on consumer hardware like Apple-silicon laptops with comparable accuracy.
Good For
- Automated Decision-Making: Ideal for applications requiring rapid, probabilistic decisions based on structured inputs.
- Routing and Classification: Can be used to route customer tickets, classify intents, or make binary choices with confidence scores.
- Resource-Constrained Environments: The efficient single-pass inference and int4 quantization make it suitable for deployment where computational resources are limited.
- Research and Development: The open-source nature of its code, data, and evaluation allows for transparent research and further development in calibrated decision models.