zils-lite-qz3
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 joamCriff123/zils-lite-l1117a, the sole body of the 18 November 2026 d3-lite soup (Apache-2.0); credit for the start belongs to its author and to the authors of the bodies it was built from. Procedure: three LoRA (rank 32, merged) fine-tunes of that body with different seeds, one epoch each at learning rate 5e-5, were averaged uniformly; no interpolation. Training targets were soft, with vllm-sr/d3-ultra (the Apache-2.0 house model of the z3 pool) as teacher: for most written items one quarter the item's own answer and three quarters d3-ultra's answer probabilities; for the about 4,100 written arithmetic-score and sentence-similarity score items d3-ultra's probabilities alone, because their written answers proved unreliable; for the 30,045 labelled rows one half the row's own dataset label and one half d3-ultra's probabilities.
Training data (about 61,100 rows, one shuffled pass per seed): (1) about 31,100 exam-format decision items of the known task families written with the public Zils authoring recipe (scripts/author_exam.py) by Claude Opus and by a local open model, from train-side splits of the public source datasets, including about 1,100 tool-call items written from rows 90,000-93,000 of the function-call subset of glaiveai/glaive-function-calling-v2 (a dataset without splits that the exam recipe also draws from); (2) 30,045 rows of the train split of tasksource/tasksource-jev-typed-decisions from 240 source tasks whose licence metadata is unspecified, used with their labels (non-commercial sources excluded; sources named like BIG-bench tasks dropped; rows whose text equals a validation or test row dropped). 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.