--- license: apache-2.0 base_model: Qwen/Qwen3.5-9B tags: [reax, sn92, risk_assessment, decision-model, merge] --- # SN92 risk_assessment specialist (v16) Jev-compatible decision model for the REAX (Bittensor SN92) `risk_assessment` specialist track: one forward pass, softmax over the option-letter logits of the `reax-decide-prompt/1` format. ## How it was built 1. **Merge.** Per-tensor linear average of two Apache-2.0 risk-assessment models that share the Qwen/Qwen3.5-9B base: * 0.65 x our own fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (revision `c202236235762e1c871ad0ccb60c8ee5ba337b9a`), and * 0.35 x [gurujustin/sn92-ra-cand-n15ab15](https://huggingface.co/gurujustin/sn92-ra-cand-n15ab15) (revision `a4264717389a882711c034aa39f488c16e037eab`), itself a merge of [edward56/sn92-ra-9b](https://huggingface.co/edward56/sn92-ra-9b) (`5cbef35f631c8b07d7e307f431201304c3d3e7a2`) and [elgin-group/sn92-ra-9b-b](https://huggingface.co/elgin-group/sn92-ra-9b-b) (`30928e221c52b6f14ce261617bd89497f73302cf`); lineage Qwen/Qwen3.5-9B -> ironbcc/seb-9b -> elgin-group/sn92-ra-9b -> edward56/sn92-ra-9b. Credit for the behaviour learned in those models goes to their authors. 2. **Fine-tuning** (LoRA, merged): REAX SN92 training rows (incl. the 17 Oct factory rows), the same rows translated into 16 languages, counterfactual variants of those rows (numbers changed, answers re-derived by two independent labeler models), synthetic decision items, and decision items computed from public data sources (see NOTICE). 3. **Distillation:** the merged-and-tuned model was further trained on the averaged answer probabilities of four of our tuned variants (an ensemble), on about 29k risk-assessment items. 4. **Readout scale:** the lm_head rows of the option letters A-Z multiplied by 2.4. Licence: Apache-2.0 (see LICENSE and NOTICE).