CharlieBT/gx33-decider

Generalist decision model (REAX SN92, m5As75)

A merge of four MIT-licensed fine-tunes of baselabs/mammoth-1 (itself built on Qwen/Qwen3.5-27B, Apache-2.0), fine-tuned to imitate a cross-family ensemble of itself and a Gemma-4-31B decision model. Same protocol as mammoth-1: joint inline readout, readout.safetensors + decision_config.json.

Lineage

parent	revision	licence	TIES weight
danyloooah/sn92-generalist-m1-core2	c78c0a4aa874d1fa87d3d2b2efe84cc19bf52c24	MIT	0.25
Lucas989/sn92-generalist-m1-core7a	2653c7d9ff002887ade9af3b2f5b65da89121e24	MIT	0.25
mufasa1228/sn92-gen-r3	c18537c947fade7acfe2beca0b130b8f73c527e3	MIT	0.25
Lucas989/sn92-generalist-m1-wide7	bb253db8d80ebfccb3cb5d3937ab188929b60908	MIT	0.25

Recipe: TIES merge (density 0.5, equal weights 0.25, sign election, lambda 1.0, readout heads averaged) of the four parents below on baselabs/mammoth-1@03007ab3c4481bf01c71ef63adbbbc4b2093d71a converted to the joint-readout layout (Qwen3_5Model backbone + readout.safetensors = the LM-head rows of the answer codes), then a full fine-tune of every weight (fp32 master weights, stored in bf16; CE + Brier on the joint readout) on ensemble-distillation targets. Released checkpoint: runs/m5A/step75.

Training data

• training targets: on 99,763 decision-task inputs (public benchmark TRAIN splits via tasksource/tasksource-jev-typed-decisions, tasksource/procedural-typed-decisions, decisionmodels/sn92-specialists and sn92-generalist-train-v1 (open-weight or programmatic authors only), nvidia/When2Call, and our own v5.1-format generators and re-labelled public records; per-source licences in TRAINING_SOURCES.tsv; no OpenAI- or Anthropic-written rows or labels) the answer distribution of a two-model logit ensemble at temperature 2: (a) tattrongvu/Jeffan-31b-1710@cebfa86c2bed2b5c83b12e44de4374b20c87fd79 (Apache-2.0, a Gemma-4-31B model; its answer probabilities only, never its weights) and (b) the TIES merge above before fine-tuning. Half of our 56 proxy task types and their source datasets are held out from training on purpose (evaluation only)

Data under CC-BY-4.0 is used with attribution to its authors (see TRAINING_SOURCES.tsv). Copyright lines of every parent are kept in LICENSE; each parent's notice is kept in NOTICE.