zils-lite-decision-v7 (malkolm777)
A d3-lite decision body (Qwen/Qwen3.5-0.8B family) for the Zils subnet 127 lite pool.
Fine-tuned from the current d3-lite soup body qhoenix/zils-lite-decision-v9 (Apache-2.0) with LoRA (rank 16, AdamW lr 2e-5) on train-split items of public reasoning and language datasets (cais/mmlu auxiliary_train, google/boolq, openai/gsm8k, facebook/anli, nyu-mll/glue rte/stsb) converted to the Zils decision format, plus the public Zils seed exams v1/v4; the fine-tune delta is scaled by 1.5. See LICENSE and NOTICE.