sn92-gen-v6

Generalist decision model for the REAX subnet (Bittensor SN92). It is a fine-tune of decisionmodels/reax-1.1 at revision 142b19355af00251f08db3d287307dd8b711c133, which Decision Models (the REAX subnet) fine-tuned from decisionmodels/reax-1 at revision ca2a0cc54a537c1f5cf24e5b09387cfb357b6518, itself a fine-tune of perplexity-ai/pplx-decider-v1.2 at revision cce0f3f80033e9d89d80831d5281fd72124aacce, which is derived from Qwen/Qwen3.8-27B. It keeps that model's protocol-v3 joint inline readout (joint_config.json, decision_config.json, readout.safetensors) and its configuration and tokenizer files unchanged. The readout head and decision_config.json are the base's.

How it was made

• Credit: the base model REAX 1.1 and the training set sn92-generalist-train-v1 are the work of Decision Models (the REAX subnet), released under Apache-2.0 and CC BY 4.0. This model is REAX 1.1 plus four LoRA deltas of ours, dW(g6), dW(t1), dW(v1t1b) and dW(x1).
• Parent entry: dW(g6) continues the LoRA training of our earlier SN92 entry Andrew93110/sn92-gen-v1 at revision 4d2a519d78377e63c46cbf35c26b09a71c28adf8 (UID 116: pplx-decider-v1.2 + dW(g4c), head x 3), so we declare that entry as this model's parent.
• dW(g6) was trained on perplexity-ai/pplx-decider-v1.2 at revision cce0f3f80033e9d89d80831d5281fd72124aacce (Apache-2.0; the model REAX 1 and REAX 1.1 descend from) and added to REAX 1.1 by task arithmetic. dW(t1) was trained on REAX 1.1 + dW(g6) with the trainer of a training kit shared with us by a partner team. All training is LoRA on the language model's attention, linear-attention and MLP projections; the readout head was never trained.
• Merge: W = W(REAX 1.1) + dW(g6) + dW(t1) + dW(v1t1b) + dW(x1), each step computed in fp32 and stored in bf16; no final-norm sharpening.
• Earlier entries these weights continue: dW(g6) continues the LoRA of our SN92 entry Andrew93110/sn92-gen-v1 at revision 4d2a519d78377e63c46cbf35c26b09a71c28adf8 (UID 116), and REAX 1.1 + dW(g6) + dW(t1) are the weights of our entry Andrew93110/sn92-gen-v3 at revision 4c37ed6ff32ac2d1d12c345772ec2eec9ada6966 (UID 71's earlier entry) before its x5 final-norm sharpening. This model adds dW(v1t1b) and is not sharpened.
• dW(v1t1b): LoRA rank 32 (alpha 64) trained on REAX 1.1 + dW(g6) + dW(t1) with the readout frozen, for the generalist evaluation v5.1 (upstream reax-subnet a0310f9 format): 5,985 items of our own v5-format data (at most 300 per task type; 11,394 rows, one per question, rendered exactly as v5.render_question serves them, 'I don't know' as the extra last option where offered) plus a replay of 2,000 sn92-generalist-train-v1 items, about half of the rows in the native decider format and half through the REAX prompt (R2 trainer), learning rate 5e-05, KL 0.3 to the start's native distribution and 0.1 to its REAX-prompt distribution, 210 steps of 64 rows (5,717,078 tokens); the adapter at the end of that pass, added at weight 1.0.
• The v5 data (ours: generators and converters in the reax-subnet ops tree, built 2026-10-25; split by family and source into training, selection and confirmation sets; the selection and confirmation sets were never trained on), traced from the rows this pass trained on:
  • openfda_device_recall_public_domain (300 items): U.S. federal government work; public domain; openFDA terms of service; U.S. Food and Drug Administration, openFDA device recall database https://open.fda.gov/terms/
  • openfda_enforcement_reports_public_domain (300 items): U.S. federal government work; public domain; openFDA terms of service; U.S. Food and Drug Administration, openFDA enforcement reports https://open.fda.gov/terms/
  • v5n_gen_rates_income_tax (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rates_parcel_postage (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rates_parking_fee (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rates_telecom_overage (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rules_data_retention (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rules_leave_request (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rules_library_loans (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5n_gen_rules_procurement_approval (300 items): Original generated items; CC0 1.0 https://creativecommons.org/publicdomain/zero/1.0/
  • v5t_procedural_chord_construction (300 items): CC0-1.0 (own computed construction)
  • xmpp_xeplist_xsf_permissive (300 items): XSF permissive licence (XEP-0001 Legal Notices: permission to copy, modify and distribute); XMPP Standards Foundation, XEP list https://xmpp.org/about/xsf/ipr-policy/
  • arxiv_api_metadata_cc0 (240 items): CC0-1.0 (arXiv API metadata: titles, abstracts, categories); arXiv API metadata, CC0 1.0 (https://info.arxiv.org/help/api/tou.html)
  • v5t_publicgen_arithmetic (240 items): CC0-1.0 (own generator)
  • v5t_publicgen_fictional_api (240 items): CC0-1.0 (own word lists and templates)
  • v5t_publicgen_logic (240 items): CC0-1.0 (own generator)
  • aegis2_cc_by_4_real_responses (200 items): CC BY 4.0 (Aegis 2.0 / Nemotron Content Safety V2); DEV-SET SOURCE: training only; Ghosh et al. 2025, AEGIS2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails, CC BY 4.0
  • hwu64_cc_by_4 (200 items): CC BY 4.0 (NLU-Evaluation-Data, Liu et al. 2019); Liu, Eshghi, Swietojanski, Rieser 2019, Benchmarking Natural Language Understanding Services for building Conversational Agents (HWU64), CC BY 4.0
  • wanli_cc_by_4 (200 items): CC BY 4.0 (WANLI); Liu, Swayamdipta, Smith, Choi 2022, WANLI: Worker and AI Collaboration for NLI Dataset Creation, CC BY 4.0
  • bitcoin_bips_per_bip_permissive (132 items): Per-BIP licence recorded in meta.licence (BSD-2-Clause, BSD-3-Clause, CC0-1.0, MIT, PD, CC-BY-4.0, OPUBL-1.0 ...); Bitcoin Improvement Proposals repository https://github.com/bitcoin/bips
  • wikidata_sparql_cc0 (124 items): CC0-1.0 (Wikidata); Wikidata (https://www.wikidata.org), CC0 1.0
  • v5t_publicgen_code (119 items): CC0-1.0 (own generator)
  • clinc150_cc_by_3 (100 items): CC BY 3.0 (CLINC150 / clinc_oos); Larson et al. 2019, An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction (CLINC150), CC BY 3.0
  • v5t_publicgen_causal (100 items): CC0-1.0 (own generator)
  • standard_ebooks_cc0_poetry_james_joyce (30 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_poetry_james_mcintyre (30 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_poetry_james_weldon_johnson (30 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_charlotte_bronte_jane_eyre (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_edith_wharton_the_house_of_mirth (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_elizabeth_gaskell_cranford (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_george_eliot_middlemarch (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_h_g_wells_the_war_of_the_worlds (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_joseph_conrad_heart_of_darkness (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_louisa_may_alcott_little_women (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
  • standard_ebooks_cc0_robert_louis_stevenson_treasure_island (20 items): CC0-1.0 edition (Standard Ebooks); source text public domain in the United States; Standard Ebooks (https://standardebooks.org), CC0 1.0 editions
• These weights also continue our entry Andrew93110/sn92-gen-v6 at revision b85f936999118e713504811440895aaa6c4265bd (REAX 1.1 + dW(g6) + dW(t1) + dW(v1t1b), committed on UID 71 and replaced by this model before it was evaluated). This model adds dW(x1) and is not sharpened.
• dW(x1): LoRA rank 32 (alpha 64) trained on REAX 1.1 + dW(g6) + dW(t1) + dW(v1t1b) with the readout frozen, for the generalist evaluation v5.1 (upstream reax-subnet a0310f9 format): 17,578 items of our own v5-format data of the computed same-skill task types (arithmetic word problems, multi-step logic, code output, chord construction, causal yes / no questions, tool-use decisions) and chess best-move questions (positions from random playouts; the gold move is Stockfish 16's choice, confirmed by a second, deeper analysis), from generator families b and c, which none of the sets we decide on uses (an earlier confirmation set of ours used family b and is retired as independent evidence), at most 2,500 per task type (18,987 rows, one per question, rendered exactly as v5.render_question serves them), plus a replay of 2,000 sn92-generalist-train-v1 items, about half of the rows in the native decider format and half through the REAX prompt (R2 trainer), learning rate 5e-05, KL 0.3 to the start's native distribution and 0.1 to its REAX-prompt distribution (the start, CB, is the teacher of both), 257 steps of 64 rows (5,513,671 tokens); the adapter at the end of that pass, added at weight 1.0.
• The x1 data (ours: generators in the reax-subnet ops tree, built 2026-10-25 from generator families b and c, which none of the sets we decide on uses; an earlier confirmation set of ours used family b and is retired as independent evidence; chess positions from random playouts, deduplicated against the public dev set's positions), traced from the rows this pass trained on:
  • v5t_publicgen_arithmetic (5000 items): CC0-1.0 (own generator)
  • v5t_procedural_random_playout_stockfish_gold (2500 items): CC0-1.0 (own generator: positions from random playouts; gold moves from Stockfish 16 analysis output)
  • v5t_publicgen_logic (2500 items): CC0-1.0 (own generator)
  • v5t_publicgen_fictional_api (2200 items): CC0-1.0 (own word lists and templates)
  • v5t_publicgen_code (1940 items): CC0-1.0 (own generator)
  • v5t_procedural_chord_construction (1938 items): CC0-1.0 (own computed construction)
  • v5t_publicgen_causal (1500 items): CC0-1.0 (own generator)

Data

• decisionmodels/sn92-specialists train rows (revision e2ff3bb099553fd9fd93f806152d5f261579bf3e, CC BY 4.0), 20 categories. Their states are disjoint from the dev split.
• Anchors, to keep the base's general knowledge and arithmetic (self-distillation; the target is the answer distribution of perplexity-ai/pplx-decider-v1.2 (the model g4 and g4c were trained on)): questions of the MMLU auxiliary_train split (cais/mmlu, revision c30699e8356da336a370243923dbaf21066bb9fe, MIT) and problems of the AQuA-RAT train split (deepmind/aqua_rat, revision 33301c6a050c96af81f63cad5562cb5363e88971, Apache-2.0).
• Area rows (gold labels), for areas the sn92 categories do not cover: math & dates (our own procedural generator: date differences, weekdays, business days, deadlines, time zones, percent change, unit conversion); tool use (glaiveai/glaive-function-calling-v2, revision e7f4b6456019f5d8bcb991ef0dd67d8ff23221ac, Apache-2.0: which function the assistant called first, and whether it called one); language understanding (MASSIVE 1.1 by Amazon, CC BY 4.0: utterance intents drawn from its 51 locales, rewritten as multiple-choice questions); knowledge (tau/commonsense_qa, revision 94630fe30dad47192a8546eb75f094926d47e155, MIT).
• decisionmodels/sn92-generalist-train-v1 (revision 2241ec2c5e9812178c3127165d830a3bc8006685, CC BY 4.0; each item is CC BY 4.0 or CC0 1.0), the REAX subnet's public generalist training set by Decision Models. Its LICENCES.md lists the upstream sources: CUAD (CC BY 4.0), Wikidata (CC0), UCI Bike Sharing (CC BY 4.0), OpenAlex (CC0) and earlier SN92 specialist data.
• Procedural decisions: the train split of tasksource/procedural-typed-decisions (revision d24af62acf13126c0a2ff548133167b838a974c6, Apache-2.0), converted to REAX questions by our code. The test split was used for evaluation only.
• Public mix A (g6): train splits of ten public datasets, converted by our code into REAX choice, yes/no and score questions with shuffled option order: tau/commonsense_qa (94630fe30dad47192a8546eb75f094926d47e155, MIT); allenai/cosmos_qa (ed50fe83db62b355759053b1cd5eef344ee371e0, CC BY 4.0); allenai/social_i_qa (537a2ec8ec565adc0b70b70752893e59e024df26, CC BY 4.0 per its card text); allenai/qasc (a34ba204eb9a33b919c10cc08f4f1c8dae5ec070, CC BY 4.0); allenai/winogrande (01e74176c63542e6b0bcb004dcdea22d94fb67b5, config winogrande_xl, CC BY per the upstream repository README (the card states no licence)); ChilleD/StrategyQA (705562638fe1d8ca6bb98c66fc8f94d45fda8c83, MIT); openai/gsm8k (740312add88f781978c0658806c59bc2815b9866, MIT); allenai/math_qa (fafb9f7ee5b9ec4da9499f9c4177a4c91389f2d6, Apache-2.0); next-tat/TAT-QA (c96247f5077eac447f63527fd3dcfdc58bb56d6a, CC BY 4.0); zeroshot/twitter-financial-news-sentiment (3e6bbcfb0c378a3017cea4ba198a4ce2594cd4c1, MIT).
• Public mix B (g6), the same conversion, for multilingual, intent, safety, law and tool-use decisions: mteb/amazon_massive_intent (dfe509a4cff77fcdead03f9275b6c0c346d049cb, the Apache-2.0 copy of Amazon's MASSIVE 1.1, CC BY 4.0; 26 non-English locales); google-research-datasets/paws-x (4cd8187c404bda33cb1f62b49b001115862acf37, PAWS-X's own licence: "may be freely used for any purpose"; 7 languages); cambridgeltl/xcopa (042f78955ba48e6404616762fa6e05e839c3907a, CC BY 4.0; the validation split in 11 languages, as XCOPA has no train split); mteb/banking77 (18072d2685ea682290f7b8924d94c62acc19c0b2, BANKING77 by PolyAI, CC BY 4.0, through the MTEB mirror whose card says MIT); clinc/clinc_oos (155b9c710419136e17307b80d0a13e68cd46b4ec, config plus, CC BY 3.0); allenai/prosocial-dialog (d77d7ad3c624c51030f2f32c83e892b3d620b3d4, CC BY 4.0); deepset/prompt-injections (4f61ecb038e9c3fb77e21034b22511b523772cdd, Apache-2.0 per its card, CC BY 4.0 per its dataset info); nvidia/Aegis-AI-Content-Safety-Dataset-2.0 (d86bb8bedff51d25ac834ab7838f1cc61acb7a2c, CC BY 4.0; human labels only); coastalcph/lex_glue (c23fdff1a6bf74e0e1a71cb86f1e781d37da888c, config case_hold, CC BY 4.0); Team-ACE/ToolACE (6bda777c88d21e5a204703c1ee45597a8fa4f734, Apache-2.0). Revisions are exact commits; for allenai/cosmos_qa, allenai/social_i_qa, allenai/math_qa, zeroshot/twitter-financial-news-sentiment and mteb/amazon_massive_intent they are commits of the refs/convert/parquet branch.
• Machine translations (g6): further sn92-specialists train rows with a text state (disjoint from g6's own sn92 rows and from the dev split) translated into Polish, Portuguese, Spanish, German, Japanese, Italian, French, Arabic, Chinese, Korean, Turkish, Dutch, Indonesian, Ukrainian and Czech with google/madlad400-3b-mt (revision fa184c675da0b5c9e1c8694fccd4e12e2d422094, Apache-2.0), greedy decoding; rows that failed the quality guards were dropped; labels and option keys unchanged.
• Our v5-format training data from 35 sources (listed in the recipe above).
• Our x1 training data from 7 sources (listed in the recipe above).

Licence

Apache-2.0 (LICENSE). NOTICE carries the base model's notice and the data attributions.