zils-mini-m2r

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 the two members of the 2026-11-11 z3-mini soup, qhoenix/zils-mini-decision-l13_avg678 (rev 091c6ebe) and joamCriff123/zils-mini-sx3 (rev 2a84fc71), both Apache-2.0. Recipe: two full-weight fine-tunes of the upper language layers (16 and above; AdamW lr 3e-6, 1.2 h each, seeds 711/712) and two LoRA members (rank 32, KL anchor 0.1 to the start, lr 5e-5, one epoch, seeds 21/22) from that start; the four members are averaged uniformly and the result is extrapolated 1.25x away from the start (1.25 x average - 0.25 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 (used by all members) and 39,636 written by a local open model, Qwen3.8-27B (heavy members only). 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.