sn92-gen-v5

Generalist decision model for the REAX subnet (Bittensor SN92): a weight average of two decision models built on Qwen/Qwen3.8-27B (declared base, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0: the common root of both and the arithmetic base of the average).

Recipe

• W = 0.5 x W(t6 (ours)) + 0.5 x W(isomsom/sn92-gen-v4) for every one of the 1,183 backbone tensors, computed in fp32 and stored in bf16.
• t6 (ours): 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, plus LoRA deltas of ours:
  • 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. t6 is REAX 1.1 plus three LoRA deltas of ours, dW(g6), dW(t1) and dW(t6).
  • Lineage of our side: 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), and REAX 1.1 + dW(g6) + dW(t1) are the weights of our entry Andrew93110/sn92-gen-v3 at 4c37ed6ff32ac2d1d12c345772ec2eec9ada6966 (UID 71, which carried a x5 sharpening in its final norm); t6 adds dW(t6) to them.
  • 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), and dW(t6) on REAX 1.1 + dW(g6) + dW(t1), 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.
  • Our side: W(t6) = W(REAX 1.1) + dW(g6) + dW(t1) + dW(t6), each step computed in fp32 and stored in bf16 (its final norm unscaled).
• isomsom/sn92-gen-v4 at revision 78cd1a860a3cb08fea2686cd057e53a91ff1d7f8: the SN92 generalist entry of isomsom (UID 140). Its card describes it as vllm-sr/d3 at revision cf15549a78fc52598ddd1951f321c58372bbd64f (Apache-2.0, a pplx-format decision model on Qwen/Qwen3.8-27B) rewritten to the joint readout layout and fine-tuned with LoRA in joint mode in three stages, with its readout rows A..Z retrained for the fixed T = 2 readout.
• Readout head: REAX 1.1's readout.safetensors, unchanged.
• Sharpening: The REAX readout is made 4 times sharper in the final RMSNorm: language_model.norm.weight stores 4 x (1 + w) - 1 of the averaged w (Qwen3.5's RMSNorm multiplies by 1 + w), and decision_config.json stores REAX 1.1's temperature x 4 (3.1606 -> 12.6424), so the native protocol's probabilities are unchanged up to numerical rounding.
• Configuration, tokenizer and joint_config.json are reax-1.1's, unchanged.

Credit

Half of this model's backbone is the work of isomsom (isomsom/sn92-gen-v4, Apache-2.0), built on vllm-sr/d3 by the vLLM Semantic Router Team (Apache-2.0). The other half of the backbone and the readout head are REAX 1.1 by Decision Models (the head unchanged) plus our LoRA deltas, which continue our earlier SN92 entries Andrew93110/sn92-gen-v1 at 4d2a519d78377e63c46cbf35c26b09a71c28adf8 (UID 116) and Andrew93110/sn92-gen-v3 at 4c37ed6ff32ac2d1d12c345772ec2eec9ada6966 (UID 71, not better on its first draw).

Selection

Read through the release's joint runtime on the public OpenDecisions validation set (jgeuter/OpenDecisions at ff0b640be2b72c7d06d9d656a34ad6eed79ec7fe, CC BY-SA 4.0, evaluation only, never trained on) for our weight 0.3, 0.35, 0.4, 0.5 and 0.65 and three readout heads (REAX 1.1's, isomsom's, and their scale-matched average) at sharpening factors 1 to 8. The weights 0.3 to 0.5 were level on OpenDecisions (within 0.3 points); 0.5 keeps the larger share of our own weights and had the best simulated pass rate under the published decision rules. isomsom's head alone reads the native decider protocol far worse than REAX 1 (top-1 on our Decision-Index proxy 65.7 vs 82.8), so it was not used; REAX 1.1's own head at factor 4 was level on OpenDecisions and keeps the native protocol close to REAX 1. A plain average of our model with vllm-sr/d3 itself did not improve on our model.

Data

Ours (t6 (ours)):

• 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.

isomsom/sn92-gen-v4:

• As stated on its card: decisionmodels/sn92-specialists train, factory and dev rows (CC BY 4.0), train splits of public benchmarks, and its authors' own written items.

Licence

Apache-2.0 (LICENSE). NOTICE carries the notices of both ingredients and the data attributions.