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. Three LoRA students (rank 64, lr 5e-5, one epoch, seeds 553 / 771 / 991) were trained from that start on the same 25,464 decision items and then weight-averaged (equal weights). Training items: exam-style decision items authored and verified by an LLM from train-split rows of public datasets (mmlu, boolq, gsm8k, anli, glue, banking77, clinc150, When2Call, poem_sentiment, iSarcasm, sarcasm headlines, SVAMP, commonsense_qa, tweet_eval, 20newsgroups, glaive function calling, BIG-bench tasks) and rows of the same public datasets in the pool's decision format; targets are soft: 0.7 x the answer distribution of our own d3-mini fine-tune (teacher) + 0.3 x the gold label. 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