putty77/sn92-generalist-c

SN92 generalist decision model (v3)

A fine-tune of perplexity-ai/pplx-decider-v1.2 (revision cce0f3f80033e9d89d80831d5281fd72124aacce; itself derived from Qwen/Qwen3.8-27B) for the REAX (Bittensor SN92) generalist track. Same architecture and request protocol as the base: Qwen3.5-family backbone with non-causal full attention, one readout per question at the <|fim_suffix|> marker (protocol v3, joint_config.json), answer probabilities = softmax(readout logits / temperature).

Changes from the base

• Fine-tuning (stages 1-7): successive LoRA stages (rank 16 each, merged into the backbone weights) plus the readout head, trained with soft cross-entropy on decision items in the SN92 schema: the REAX / decisionmodels SN92 specialists dataset (CC BY 4.0, train split of all 20 categories and part of the dev split), rows of all categories machine-translated into 16 languages, counterfactual variants of risk-assessment rows (numbers changed, answers re-derived by two independent labeler models), programmatically generated math & dates decision items with exact answers, general-knowledge multiple-choice questions (MMLU auxiliary training set: ARC, OpenBookQA, RACE, MCTest), XNLI natural-language-inference items in 10 languages, and decision items computed from public data sources (attributions in NOTICE). Stage 7 is an equal-weight average (model soup) of two variants trained from the same parent (one on the Decider joint protocol, one on a mix of that protocol and the SN92 REAX single-question prompt).
• Stages 8-9: two further LoRA stages (rank 32 each) on the REAX / decisionmodels SN92 generalist training data (train v1, CC BY 4.0; a category-balanced sample of about 53k items in 27 languages, then about 24k further items of that set) plus replay of the earlier data, with 70 % of the rows in the SN92 REAX single-question prompt format (readout at the last prompt position) and 30 % in the Decider joint protocol.
• Temperature: the refit answer temperature is folded into the readout head (head scaled by 6.0) and decision_config.json says 2.0, so the native Decider path and a fixed-temperature-2 readout give the same probabilities.

All credit for the base model goes to Perplexity. Licence: Apache-2.0 (see LICENSE and NOTICE).