zils-lite-jq6
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.
Soup fine-tune: the start is the uniform weight average of the two bodies of the 14 November 2026 d3-lite soup (VadymPikulyk/zils-lite-dh2 and qhoenix/zils-lite-decision-avg-abc), both Apache-2.0; credit for the start belongs to their authors and to the authors of the bodies they were built from. Procedure: six LoRA (rank 32, merged) fine-tunes of that average with different seeds, one pass each at learning rate 5e-5 (three over 30,000 rows and three over 45,913 rows of the data below), were averaged uniformly with equal weights. Training targets were soft and label-free: the answer probabilities of the current d3-mini soup average (teacher distillation from the larger pool's soup: the uniform weight average of Zayaan/zils-mini-hm4, qhoenix/zils-mini-decision-t2avg and joamCriff123/zils-mini-m1113b).
Training data (30,000 or 45,913 rows per seed, one shuffled pass): rows of the train split of tasksource/tasksource-jev-typed-decisions (about 630 source tasks; sources with commercial or unspecified licence metadata only, non-commercial sources excluded; sources whose names match BIG-bench tasks dropped; no Glaive rows), used only as unlabeled distillation inputs (their labels were not used). No exam-format written items, no validator exam items, no public seed exam items and no customer data were used. This repository holds weights only; check the source datasets' terms before reuse.