Zils d3-lite decision body
A fine-tune of Qwen/Qwen3.5-0.8B in the chassis layout for the Zils d3-lite soup pool (Bittensor subnet 127).
Start: the uniform average of the current soup members Charlie320/zils-lite-v9a and qhoenix/zils-lite-decision-v9 (Apache-2.0).
Method: several one-epoch LoRA fine-tunes (official zils.teacher train, ranks 16-64) with different seeds, merged and averaged.
Training data (train splits only, about 88k rows): soft answers from zils-ai/zils-relay-answers; tasksource/tasksource-jev-typed-decisions (train, commercial-use sources); decisionmodels/sn92-specialists and decisionmodels/sn92-generalist-train-v1 (CC-BY-4.0); samatv256/jev-decisions-v1 (CC-BY-4.0); exam-format items written with the Zils authoring recipe; anchor rows answered by the start soup. Rows matching public benchmark test sets were removed. Check each source's licence before reuse.