zils-mini-mx8
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: the start is joamCriff123/zils-mini-m1115, the sole body of the 18 November 2026 d3-mini soup (Apache-2.0); credit for the start belongs to its author and to the authors of the bodies it was built from. Procedure: two LoRA (rank 64, merged) fine-tunes of that body with different seeds, one epoch each at learning rate 1e-4, were averaged uniformly; no interpolation. Training targets were soft: three quarters the row's own dataset label and one quarter the starting body's own answer probabilities (self-anchoring against forgetting); no other model was used as a teacher.
Training data (30,000 rows, one shuffled pass per seed): rows of the train split of tasksource/tasksource-jev-typed-decisions from 240 source tasks whose licence metadata is unspecified (non-commercial sources excluded; sources named like BIG-bench tasks dropped; rows whose text equals a validation or test row dropped), used with their labels. No written exam-format 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.