zils-lite-body-3

Zils d3-lite (0.8B) soup-pool body: LoRA fine-tune (merged, bf16) starting from the uniform average of the current lite soup bodies, trained on exam-style decisions built from public training splits (NLI, similarity, QA, intents, topic, sentiment, arithmetic), computed tool-call / counting fixtures, UniMorph verb forms, headline sarcasm, Stanford politeness, real-world outcome events (zils.current) and Zils Relay teacher probabilities, plus self-distillation regularisation. Vision tower unchanged. Apache-2.0; upstream data keeps its own licences.