johnmiller31109/sn92-ra-9b

Specialist decision model for the REAX (Bittensor subnet 92) risk_assessment track. It answers typed decisions (choice, yes/no, ordered score) in the reax-decide-prompt/1 format: one forward pass, softmax over the option-letter logits at the first generated position (temperature 2, thinking disabled).

• Base: Qwen/Qwen3.5-9B@c202236235762e1c871ad0ccb60c8ee5ba337b9a (Apache-2.0)
• Weights: bf16 safetensors, text-only Qwen3_5ForCausalLM
• Licence: Apache-2.0 (see LICENSE and NOTICE)

Training data (licences as declared by each source):

• programmatically generated typed decisions (original, Apache-2.0)
• tasksource/procedural-typed-decisions (Apache-2.0)
• totum-labs/plumb-data, kev_hard (Apache-2.0)
• crh225/plumb-decisions (Apache-2.0)
• LocalLLaMA/typed-decisions, train split (Apache-2.0)
• tasksource/tasksource-jev-typed-decisions, rows marked for commercial use (CC-BY-4.0 / MIT / Apache-2.0 / CDLA-Permissive-1.0 per row; CC-BY-4.0 sources are attributed to their original authors via tasksource)