CharlieBT/gx27-decider

Generalist decision model (REAX SN92, df55)

A LoRA fine-tune of CharlieBT/gx26-decider (Apache-2.0), merged into the weights, for the REAX subnet's generalist decision track. Same protocol as the base: joint inline-readout v3, readout.safetensors + decision_config.json.

Training data (see MODIFICATIONS.md for the full list):

• our own v5.1-format task generators and re-labelled public records (per-row source + licence in TRAINING_SOURCES.tsv): GitHub Advisory Database (CC-BY-4.0), TED EU tender notices (European Commission reuse notice, (c) European Union), English Wikipedia deletion discussions (CC-BY-SA-4.0; page title + revision id per row), openFDA / CPSC / NWS METAR records (US government, public domain), OurAirports and openfootball (public domain), arXiv metadata and Wikidata (CC0), Ethereum EIPs (CC0) and other improvement-proposal registries (Apache-2.0 / BSD / CC-BY / CC0 / MIT, per row), Project Gutenberg texts (public domain in the US), SNLI / MultiNLI (CC-BY-SA / OANC), ANLI TRAIN split (CC-BY-NC-4.0), IBM Argument Quality 30k (CC-BY-SA-3.0), Amazon MASSIVE (CC-BY-4.0), NVIDIA Aegis 2.0 (CC-BY-4.0); procedural generators (CC0)
• decisionmodels/sn92-generalist-train-v1 @ 2241ec2c (REAX / decisionmodels, CC-BY-4.0): only rows written by open-weight models or by code (author families qwen, deepseek, zai-org, google, moonshot, code, programmatic, sn92-procgen); no OpenAI- or Anthropic-written rows
• decisionmodels/sn92-specialists (REAX / decisionmodels, CC-BY-4.0): train split rows
• nvidia/When2Call @ 0582f774 (NVIDIA, CC-BY-4.0; generated with Mixtral-8x22B-Instruct): TRAIN preference split only, decontaminated against its test sets
• public benchmark TRAIN splits converted to the decision format (each under its own licence; list in TRAINING_SOURCES.tsv); no Decision-Index test-derived source
• decision items written for this model by open-weight models (Qwen, DeepSeek, GLM, Kimi) and by our own programmatic generators (dates, calendars, finance, policy rules, code/SQL, tool routing, abstention pairs)
• replay rows: a sample of the rows above re-labelled with the parent's own served distribution (CharlieBT/gx26-decider@199465b0, softmax of its logits at T = 2) as soft targets, i.e. self-distillation toward the parent against forgetting
• format-varied copies of training items (second stage): question and header wording from paraphrase templates written by Kimi-K3 (open-weight) via the engy gateway; record layouts, board notations and option sets re-rendered by code; same records, labels and licences as the originals

Data under CC-BY-4.0 is used with attribution to its authors as listed above.