sn92-generalist-b1

SN92 (REAX) generalist entry (eval v5.1). Fine-tune of baselabs/mammoth-1 @ 03007ab3c4481bf01c71ef63adbbbc4b2093d71a (post-trained from Qwen/Qwen3.5-27B), in the REAX joint-readout layout (Qwen3_5Model backbone + readout head): 1 LoRA run merged into the weights plus the fine-tuned readout head, trained on procedurally generated items with computed gold (arithmetic word problems, multi-step logic, Python code output, chord construction, chess positions with Stockfish gold, and 15 rule-application families: tariffs, policies, fees, eligibility, scheduling), 600 safety-labelling rows built from NVIDIA Aegis 2.0 prompts (CC BY 4.0) with fixed refusal templates, 1,555 document-to-official-category items built from public labelled records (patent abstracts with CPC sections from ccdv/patent-classification; LEDGAR contract provisions, Supreme Court issue areas and CaseHOLD holdings from coastalcph/lex_glue; DBpedia ontology classes from fancyzhx/dbpedia_14; condition categories from TimSchopf/medical_abstracts; development-status and topic classifiers from the PyPI JSON API), each under its own terms, 'I don't know' rows (ARC / OpenBookQA / SuperGPQA questions that decisionmodels/reax-1.1 answered wrongly or unsurely), and rows of decisionmodels/sn92-generalist-train-v1 (CC BY 4.0). See MODIFICATIONS.md, NOTICE and LICENSE. Licence: MIT (base terms retained; Qwen3.5-27B under Apache-2.0).