misa-jsc/misa-ai-2.0-decision-plus

MISA AI 2.0 Decision Plus

MISA AI 2.0 Decision Plus is a decision model from MISA JSC. It is fine-tuned from Gemma 4 31B. It answers structured questions about a piece of text or an image with calibrated probabilities, rather than generating free text. Use it when you need a fast, reliable decision, such as:

• Classifying intent, topic, or sentiment.
• Flagging policy or safety violations.
• Routing a message to the right agent or tool.
• Scoring a severity or urgency level.

Model details

Developer	MISA JSC
Architecture	Gemma4ForConditionalGeneration, dense, 60 layers, hidden size 5376
Parameters	~31B
Context length	256K tokens
Vocabulary	262K
Inputs	Text, images
Output	Probabilities over answer labels (the model generates no tokens)
Precision	bf16 safetensors, 2 shards, ~62.6 GB

How it works

Each request has a state (the text to judge, plus optional images) and one or more questions. The model reads the state once in a single prefill pass. It then returns a probability distribution over each question's answer labels. No tokens are sampled, so results are deterministic and latency is close to prefill time.

Three question types are supported:

Type	Answer
noul (yes/no)	Probability that the statement is true
choice	One label out of a set of named options, with per-option probabilities and a confidence
score	An ordinal level from an ordered list (e.g. none / mild / severe), with per-level probabilities

Each answer also returns a label mass (x_label_mass / label_mass). This is the share of the model's probability that fell on the valid labels. A low label mass means the question did not fit the state well, so treat that answer as low-confidence.

Usage

Serving with SGLang

Recent SGLang releases serve /v1/systemone and /v1/decisions out of the box:

python3 -m sglang.launch_server \
  --model-path misa-jsc/misa-ai-2.0-decision-plus \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --port 30000

curl -X POST http://localhost:30000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{
    "model": "misa-jsc/misa-ai-2.0-decision-plus",
    "state": "Customer: I paid 3 days ago and still have not received my invoice. This is frustrating!",
    "questions": {
      "complaint": {"type": "noul", "instructions": "The customer is complaining"},
      "intent": {"type": "choice", "instructions": "Main purpose of the message",
                 "criteria": {"question": "Asking for information", "complaint": "Complaint", "order": "Placing an order"}}
    }
  }'

To pass images, add an images array of base64 data URLs or {"url": ...} objects.

Tips

• Prefer one choice question over many yes/no questions when the answers are mutually exclusive (e.g. which tool to call). It is faster and better calibrated.
• Keep instructions short and concrete. Name the subject explicitly ("The customer …", "The image …").
• Thinking mode is not used. Decisions are read directly from the first answer token.

Limitations

• The model is built for decisions over a given state. It is not a knowledge base or a chat assistant.
• Probabilities are calibrated on the training distribution. On very different domains, check calibration before relying on thresholds.
• For safety and moderation use, set thresholds per category on your own traffic. Review false positives on domain jargon (e.g. accounting or retail terms) before blocking users automatically.

License

The weights are released under the Gemma 4 license (Apache 2.0), inherited from the base model.