zils-flash-fg3
A body for the Zils d3-flash soup pool (Bittensor subnet 127): a fine-tune of Qwen/Qwen3.5-9B in the chassis layout, BF16, language weights trained, vision tower from the parent.
Fine-tune of the d3-flash house model: the start is vllm-sr/d3-flash (Apache-2.0, built on Qwen/Qwen3.5-9B); credit for the start belongs to its authors. Procedure: three LoRA (rank 32, merged) fine-tunes of that start with different seeds, one shuffled pass each over the 61,091 rows below at learning rate 5e-5, were averaged uniformly with equal weights. Training targets were soft: for the written exam-format items three quarters the item's own answer and one quarter the answer probabilities of the current d3-mini soup average (the uniform weight average of Zayaan/zils-mini-hm4, qhoenix/zils-mini-decision-t2avg and joamCriff123/zils-mini-m1113b); for the tasksource-jev rows one half the row's own label and one half that soup average's probabilities.
Training data (61,091 rows): (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) 30,000 rows of the train split of tasksource/tasksource-jev-typed-decisions with their labels (sources with commercial or unspecified licence metadata only, non-commercial sources excluded; sources whose names match BIG-bench tasks 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.