A d3-lite decision body (Qwen/Qwen3.5-0.8B family) for the Zils subnet 127 lite soup pool.
Start: a plain weight average of the two current lite pool members (VadymPikulyk/zils-lite-dh2 and qhoenix/zils-lite-decision-avg-abc), i.e. the pool's soup. Four LoRA students (rank 64, lr 5e-5, one epoch, seeds 553 / 771 / 991 / 1201) were trained from that start, weight-averaged, and the resulting change from the start was scaled by 1.5 (delta extrapolation). Training set: 4,580 exam-style decision items in the pool's task families, authored by an LLM in the validator's format (state + decision question with declared options). Targets are the full option distributions of the larger d3-family model vllm-sr/d3-ultra on each item (teacher distillation, KL); no gold labels were used. Readout, vision tower and all contract files are the chassis'. Licensed under Apache-2.0; see LICENSE and NOTICE.
model-00001-of-00001.safetensors 1.7 GBreadout.safetensors 1 MBvllm-sr/d3-lite