zils-mini-body-10: long-tail breadth LoRA fine-tune of the current Zils MINI soup

A decision body for the Zils d3-mini soup pool (Bittensor subnet 127), in the d3-mini chassis layout (Qwen3.5-4B language and vision tensors, bf16). It is a fine-tune of the current MINI soup: the uniform weight average (official zils.soup.build: bf16-canonical, fp32 mean, chassis vision tower) of qhoenix/zils-mini-decision-v2 @ecb9cadb39a2f84cefd2f27a755ec12095b62933 and Charlie320/zils-mini-s1 @53eb3ffa768c1ad1eace7acbfdb3ebb9674cefe6 (soup body digest df87aaa9…), as Zils's mining guide allows ("You may fine-tune the current soup"). Credit for the starting weights belongs to the authors of those two bodies and their upstream works (see NOTICE). No other miner's body was merged; it is not a re-upload of any public model.

Method

Training: LoRA rank 32 / alpha 64 / dropout 0.05 on the attention and Gated-DeltaNet projections of all language layers (152 matrices), learning rate 5e-5, AdamW with weight decay 1.0 on the LoRA weights only, batch 8 x accumulation 2, warm-up 50 steps then cosine, 2 epochs, max 2,048 tokens, seed 4343; loss per row = cross-entropy on the row's label + 1.0 x Brier. No KL / distillation term, no reference or teacher probabilities. Merged into bf16. This body is the final save.
Data (3,208 rows):
- 2,803 relay rows from 70 long-tail public task sources (tasksource collection of public TRAIN splits, at most 150 per family). Their labels are HOUSE labels: the argmax answers of a Zils house teacher model in the archived relay pool (attributed to d3-ultra; exact teacher revision unknown). They are house hard targets, not gold labels; Zils allows house targets.
- 405 anchor rows with gold labels: code-generated arts / language / reasoning score items (labels computed by code), GoEmotions (simplified) train rows, and MNLI / PIQA / GLUE train-split rows with their dataset labels.
- No other miner's outputs were used as targets. Every row passed our exam-split 8-gram guard (81 exam sources, 0 hits); no Glaive data; no template tools items.

Checks we ran before commit (not an official score)

Change from Qwen/Qwen3.5-4B (Zils relative change): 0.01254 (cap 0.15). Zils body digest 8a3c157a8d14770bfc0bd0d5cf6b7f02c307c8c8ee68d9f6114a0afe4792f366.
Copy screen with the official soup_copies code (our seeds; the validator's seed is private): own share beyond the house bodies, the listed public models, the soup and the release 0.114-0.123 (house-mix line 0.05); largest cosine to any of them or to our sibling bodies 0.941 (line 0.999); largest share explained by earlier bodies 0.223 (line 0.30).
Submitted on the owner's decision before our full pre-commit read completed.