zils-lite-lsd2

A d3-lite decision body (Qwen/Qwen3.5-0.8B family) for the Zils subnet 127 lite pool.

Built by LoRA fine-tuning the current d3-lite soup (qhoenix/zils-lite-decision-v9, Apache-2.0) on exam-format training items (written for this purpose with OpenAI Codex, DeepSeek and Gemini, each answer blind-checked by a second pass, plus the public Zils seed exams zils-ai/base-exam-seed-v1 and -v4) and exam-format rows converted from public dataset train splits (BIG-bench via tasksource/bigbench, SciQ, COPA, BoolQ, MRPC, QQP, ARC, OpenBookQA, CommonsenseQA, HellaSwag, WinoGrande, PAWS, ANLI and others), with each target mixed 0.2 label + 0.8 the predictions of the d3-mini soup (uniform weight average of qhoenix/zils-mini-decision-v2 and Charlie320/zils-mini-s1, both Apache-2.0) used as a distillation teacher; the body is the soup plus one such LoRA run. Licensed under Apache-2.0; see LICENSE and NOTICE for upstream attributions.