isomsom/sn92-gen-v91a

REAX SN92 generalist entry

Decision model for the REAX (SN92) generalist track, served with the validator's fixed prompt and joint letter readout (softmax at T=2).

Lineage and modifications

Qwen/Qwen3.5-27B (Apache-2.0) -> baselabs/mammoth-1@61cec44c281c693992e3d6ebeb44dd3402dfa186 (MIT; a full fine-tune for decisions by Base Labs, Baseten) -> danyloooah/sn92-generalist-m1-core2@c78c0a4aa874d1fa87d3d2b2efe84cc19bf52c24 (MIT; a merge on mammoth-1 in the joint readout layout whose published parents are isomsom/sn92-gen-m87b@7f04058f63745440fae3c10f8da249a0c55fea40 (MIT, our earlier entry: a LoRA fine-tune of mammoth-1), xflud/sn92-decider-1@db2e622aa508bd82350620bf165441fa4043f3cf (MIT) and danyloooah/sn92-generalist-m1-v3a@ac7fe8d04867802926ac9b05d16592a7b53bc26a (MIT)), its readout rows beyond Z (codes AA and later) re-derived from the model's own A..Z scale (least-squares fit of rows A..Z on mammoth-1's, applied to mammoth-1's rows beyond Z; rows A..Z unchanged) -> a NEW LoRA r32 (alpha 64, all linear layers of the language model; the vision tower untouched) trained in joint mode (bidirectional full attention over the prompt, as the REAX validator serves it) for 300 steps (of a 600-step cosine schedule) of 32 rows (lr 2.5e-5, seed 91) with a letter head over all 255 answer codes (questions with more than 26 options labelled AA, AB, ...), on decision items in the REAX item format: a sample of our m87b training data and the public SN92 specialist replay (decisionmodels/sn92-specialists, CC-BY-4.0; tasksource/procedural-typed-decisions, Apache-2.0), with every generated item (our own generators, the procedural and specialist-replay items, the tool-selection items) and a part of the rest anchored by KL divergence to the starting model's own answers (no gold label on those items); multiple-choice knowledge items from the TRAIN splits of CommonsenseQA (MIT), OpenBookQA, MathQA and AQuA-RAT (Apache-2.0); intent-classification items over the full label sets of BANKING77 and MASSIVE en-US (CC BY 4.0) and CLINC150 (CC BY 3.0), TRAIN splits; tool-selection items over fictional catalogues we wrote; multi-label yes/no items over PLOS abstracts and subject terms (CC BY 4.0), arXiv abstracts and categories (arXiv metadata, CC0) and fictional label banks we wrote; merged into the backbone; all readout rows rescaled by 1.0 and shifted by the mean of rows A and B (a common shift that leaves the option softmax unchanged) for the fixed T=2 readout.

Parents

• baselabs/mammoth-1@61cec44c281c693992e3d6ebeb44dd3402dfa186 (MIT; a fine-tune of Qwen/Qwen3.5-27B, Apache-2.0): the declared base.
• danyloooah/sn92-generalist-m1-core2@c78c0a4aa874d1fa87d3d2b2efe84cc19bf52c24 (MIT)
• isomsom/sn92-gen-m87b@7f04058f63745440fae3c10f8da249a0c55fea40 (MIT)
• xflud/sn92-decider-1@db2e622aa508bd82350620bf165441fa4043f3cf (MIT)
• danyloooah/sn92-generalist-m1-v3a@ac7fe8d04867802926ac9b05d16592a7b53bc26a (MIT)

Their licence files are reproduced in NOTICE.

Training data credits

Training data attribution. The model was trained on decision items (questions with answer options) derived from the sources below; texts were shortened, masked, re-ordered and placed into decision questions (modified from the original). TRAIN splits only wherever a dataset has splits. Licence texts: CC BY 4.0 https://creativecommons.org/licenses/by/4.0/, CC BY 3.0 https://creativecommons.org/licenses/by/3.0/, CC BY-SA 4.0 https://creativecommons.org/licenses/by-sa/4.0/, Apache-2.0 https://www.apache.org/licenses/LICENSE-2.0; the MIT licence text is in LICENSE.

Intent classification (utterances placed into decision questions over the intent labels):

• BANKING77 (PolyAI/banking77), Casanueva et al., "Efficient Intent Detection with Dual Sentence Encoders" (2020), licensed under CC BY 4.0
• MASSIVE (AmazonScience/massive, en-US), FitzGerald et al., "MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset with 51 Typologically-Diverse Languages" (2022), Amazon.com, Inc., licensed under CC BY 4.0
• CLINC150 (clinc_oos, "plus"), Larson et al., "An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction" (EMNLP-IJCNLP 2019), licensed under CC BY 3.0 Multi-label yes/no decision items (one candidate label per item):
• Stack Exchange posts (question title and body) and their tags, user contributions licensed under CC BY-SA 4.0 (Stack Exchange, Inc. and the post authors; each item derives from one question page on its Stack Exchange site)
• PLOS article titles, abstracts and PLOS thesaurus subject terms (PLOS Search API), licensed under CC BY 4.0 (Public Library of Science and the article authors)
• arXiv titles, abstracts and category listings from the arXiv API (descriptive metadata, CC0 1.0); thank you to arXiv for use of its open access interoperability Record classification and outcome items (one record or discussion per item, the label it was given hidden):
• English Wikipedia: Articles / Categories / Templates / Redirects for discussion, Requests for adminship and Requested moves discussion pages, by the Wikipedia contributors (https://en.wikipedia.org), licensed under CC BY-SA 4.0
• GitHub Advisory Database (GitHub, Inc.; https://github.com/advisories), records via the OSV.dev exports, licensed under CC BY 4.0
• Contains information licensed under the Open Government Licence - Canada (open.canada.ca dataset records; https://open.canada.ca/en/open-government-licence-canada)
• EUR-Lex directory records: (c) European Union, https://eur-lex.europa.eu; TED (Tenders Electronic Daily) contract notices: (c) European Union, https://ted.europa.eu (Publications Office of the European Union; reuse under Commission Decision 2011/833/EU)
• Contains public sector information licensed under the Open Government Licence v3.0 (legislation.gov.uk, Crown copyright)
• Contains Parliamentary information licensed under the Open Parliament Licence v3.0 (UK Parliament Written Questions)
• Art Institute of Chicago API (artic.edu): record data CC0 1.0, artwork descriptions CC BY 4.0 (Art Institute of Chicago)
• CVE records via the NVD (NIST): CVE(R) content Copyright (c) The MITRE Corporation, used under the CVE Terms of Use (https://www.cve.org/Legal/TermsOfUse); CVE is a registered trademark of The MITRE Corporation
• Public-domain U.S. government works and CC0 records: NASA Image and Video Library and NTRS, USGS Publications Warehouse, CPSC recalls, MSHA, OSHA, NOAA NCEI Storm Events, openFDA, FederalRegister.gov, NIH RePORTER, NSF awards, USDA FoodData Central; Wikidata, Project Gutenberg catalog, The Metropolitan Museum of Art Open Access, OpenAlex (CC0 1.0) Tool-use items: part of the API catalogues are real API descriptions: Stripe OpenAPI (Copyright (c) Stripe, Inc., MIT), GitHub REST API description (Copyright (c) GitHub, Inc., MIT), Twilio OpenAPI (Copyright (c) Twilio Inc., MIT), Kubernetes OpenAPI (The Kubernetes Authors, Apache-2.0), Box OpenAPI (Box, Inc., Apache-2.0); the rest are fictional. Multiple-choice knowledge items: CommonsenseQA (Talmor et al. 2019, MIT), OpenBookQA (Mihaylov et al. 2018, Apache-2.0), MathQA (Amini et al. 2019, Apache-2.0), AQuA-RAT (Ling et al. 2017, DeepMind, Apache-2.0). Decision-item datasets (rows used as published, in the REAX item format):
• decisionmodels/sn92-specialists, licensed under CC BY 4.0
• tasksource/procedural-typed-decisions, Apache-2.0
• tasksource/tasksource-jev-typed-decisions (licence per row, recorded there), with items from: CosmosQA, QASC, QuaRTz, Social IQa, TabFact, Circa, SLURP, KnowRef-60k, WinoDict (CC BY 4.0); Spider, SQuAD 2.0, Schema-Guided Dialogue, CREAK, probability_words_nli (CC BY-SA 4.0); CONDAQA, EQUATE, InfoTabs, SpartQA, Temporal NLI, TRACIE (Apache-2.0); CommonsenseQA, NumerSense, Corr2Cause (MIT); bAbI (BSD); CODAH (ODC-BY); SNIPS built-in intents (CC0 1.0)
• m-lagnajit/calibrated-typed-decisions (items from BANKING77 and MASSIVE, CC BY 4.0, and CLINC150, CC BY 3.0)
• crh225/plumb-decisions, Apache-2.0
• NousResearch/hermes-function-calling-v1, Apache-2.0
• openai/gsm8k, Copyright (c) 2021 OpenAI, MIT License
• ZefanCai/Open-Jev-v1.1, CC0 1.0 The other items were generated by us (fictional records, catalogues and rule sets, games, arithmetic, code and SQL).

Declared base for the track: baselabs/mammoth-1@61cec44c281c693992e3d6ebeb44dd3402dfa186. Weights only; no code. Licence: MIT (see LICENSE and NOTICE; the base of mammoth-1, Qwen/Qwen3.5-27B, is Apache-2.0).

Not affiliated with, endorsed by, or sponsored by Base Labs / Baseten, the Qwen team or the authors of the parent models and datasets.