zils-mini-sg3

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.

Soup fine-tune of the 11 November mini soup: three rank-64 LoRA runs (lr 5e-5, one epoch, different seeds) averaged.

Training data: ~84k decision items: ~70k converted from public TRAIN splits with their gold labels (MNLI, RTE, STS-B, 20 Newsgroups, QNLI, emotion, sarcasm, SIQA, PIQA, GSM8K, Glaive function calling, TweetEval-derived rule scores) in the current exam formats; 7,120 exam-format items written with the public Zils authoring recipe (x2); and the public seed exam zils-ai/base-exam-seed-v1 (old families, x3). Seed exams v2-v4 were not used for training.