zils-mini-hx3

A body for the Zils d3-mini soup pool (Bittensor subnet 127): a fine-tune of Qwen/Qwen3.5-4B in the chassis layout, BF16, language weights trained, vision tower from the parent.

Start: uniform weight average of three z3-mini house bodies, perplexity-ai/pplx-decider-flash-v1.2 (rev 9ff7cf98), Maincode/Matilda-Decision-v1-Mini (rev 39479a8a) and PelaAI/KnowLine-4B-Gen5 (rev a8e0cd66), all Apache-2.0. Recipe: one full-weight fine-tune of the upper language layers (16 and above; AdamW lr 4e-6, 1.2 h, seed 721) and one LoRA member (rank 64, lr 5e-5, one epoch, seed 722) from that start; the two members are averaged 50/50 and the result is extrapolated 1.3x away from the start (1.3 x average - 0.3 x start). Answer layer and vision tower unchanged.

Training data: 46,756 exam-format decision items written from public training-split rows with the public Zils authoring recipe (scripts/author_exam.py): 7,120 written by Claude Opus and 39,636 written by a local open model, Qwen3.8-27B. No validator exam items, no public seed exam items and no relay rows were used. This repository holds weights only; check the source datasets terms before reuse.