zils-lite-dh8
A z3-lite (d3-lite) soup body for Zils, Bittensor subnet 127: the language and vision weights of a fine-tune of Qwen/Qwen3.5-0.8B in the d3-lite chassis layout.
It starts from the uniform average of the two seated d3-lite bodies (VadymPikulyk/zils-lite-dh2 and
qhoenix/zils-lite-decision-avg-abc, both Apache-2.0) and adds 0.7 x a merged LoRA (rank 64, lr 5e-5, one epoch)
trained with python -m zils.teacher train on about 26,000 Zils Relay teacher answers (zils-ai/zils-relay-answers,
vllm-sr/d3-ultra probabilities on decision questions rebuilt from public training sources) plus 7,200 tool-use
decision rows built from the public TRAIN split of glaiveai/glaive-function-calling-v2, whose targets keep the
starting soup's own answers (70%) mixed with the gold label (30%). No validator exam items, public seed exam items
or customer data were used.
Apache-2.0, as its parents (see LICENSE and NOTICE). Please cite tasksource and the original datasets.