zils-mini-body-6: exact-label hard-item 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 language layers (no MLP adapters), learning rate 5e-5, AdamW with weight decay 1.0 on the LoRA weights only, batch 8 x accumulation 2, 2 epochs, 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 (1,315 rows, hard labels only): the items of our verified question bank that the starting soup answers with less than 0.90 probability on the gold option:
- 789 logic items with code-derived labels: logical-entailment (labels re-derived by full truth table) and RuleTaker (labels re-derived by a forward-chaining solver), public TRAIN splits;
- 365 code-generated score items (reasoning 157, arts 122, language 86; labels computed by code and an independent recount);
- 105 arithmetic word problems from GSM8K / SVAMP train splits, 27 HellaSwag / ARC train items and 2 CoQA train items, with their dataset labels;
- 27 tool-use items built from public function-calling records (ToolACE, Hermes function-calling).
- No house labels, no writer-model labels, no other miner's outputs. Every row passed our exam-split 8-gram guard; no Glaive data.

Checks we ran before commit (not an official score)

Change from Qwen/Qwen3.5-4B (Zils relative change): 0.01212 (cap 0.15). Zils body digest 2037214bd24bd852b2e675484b81047c6bdd82fd8fec101b12258c7f40cd4077.
Copy screen with the official soup_copies code (our seeds; the validator's seed is private): own share 0.0555-0.0607 (house-mix line 0.05); largest cosine 0.972, to the soup (line 0.999); largest share explained by earlier bodies 0.14-0.15 (line 0.30).
Submitted on the owner's decision. Our internal proxy read showed gains in language and arts and losses in reasoning and retrieval.