zils-lite-hm1
A body for the Zils d3-lite soup pool (Bittensor subnet 127): a fine-tune of Qwen/Qwen3.5-0.8B in the chassis layout, BF16, language weights trained, vision tower from the parent.
Start: the house body Maincode/Matilda-Decision-v1-Edge (Apache-2.0). Stage 1: full fine-tune of every language layer (lr 4e-6, cosine schedule, about 234k rows, cross-entropy + Brier loss, with a KL anchor to the start body on 40k broad rows). Stage 2: one epoch of LoRA (rank 32, alpha 64, lr 2e-4, merged) on 4.1k exam-format items. The body is the uniform weight average of the stage-1 and stage-2 weights.
Training data: exam-format decision items written with the public Zils authoring recipe (scripts/author_exam.py) by Claude Opus and by a local open model, from public dataset rows: train-side splits for the known task families; for the new-domain style tasks, rows of the public new-domain task pools listed in scripts/author_exam.py (BIG-bench task files and splits of other public datasets, mostly held-out validation/test splits; BIG-bench tasks used by the public seed exams were excluded), plus the public Zils relay teacher answers. No validator exam items and no public seed items were used. This repository holds weights only; check the source datasets' terms before reuse.