jwcwn/zl-mini-body-4

zl-mini-n2

A d3-mini soup body for Zils (Bittensor subnet 127).

• Start: a uniform average of qhoenix/zils-mini-decision-v2 @ ecb9cadb and Charlie320/zils-mini-s1 @ 53eb3ffa, both Apache-2.0, built in the d3-mini chassis (Qwen3.5-4B).
• Training: LoRA rank 32, lr 3e-5, one epoch, soft cross-entropy plus Brier loss with a KL anchor (0.5) to the start.
• Data: about 3,800 exam-format items for the 15 known task families. Items were written by open-weight models (Qwen3.5-35B-A3B author, gpt-oss-20b blind checker, items kept only where the checker agreed with the label) or computed by code, using only train splits of public datasets as topic hints. Items were deduplicated against held-out evaluation data by source row, normalized fingerprints, option order and embedding similarity.

Weights are bf16 safetensors in the pool's layout. License: Apache-2.0, as its parents.