zils-mini-body-2

Repository: tensor-tailor/zils-mini-body-2.

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 starts from the uniform average of three d3-mini house bodies and adds a merged-LoRA fine-tune. It is a relative of our zils-mini-body-1: same starting average and the same LoRA trainer settings. Two things differ from that published body: (a) the training data here carries no teacher-distillation rows (see "Training data"), and (b) this body is the plain merged LoRA fine-tune of the starting average, whereas zils-mini-body-1 was a further average of that fine-tune back toward the start. It is not a re-upload of any public model and does not use the weights of any other miner's body.

Method

• Starting weights: the uniform fp32 average, over all chassis tensors, of torchcast-ai/torchcast-decision-4b @492bfb19 (Apache-2.0), Maincode/Matilda-Decision-v1-Mini @39479a8a (Apache-2.0) and perplexity-ai/pplx-decider-flash-v1.2 @9ff7cf98 (Apache-2.0). The chassis is zils-ai/d3-mini-4.5b-v1 @b54402a9 (Apache-2.0); parent Qwen/Qwen3.5-4B @851bf6e8 (Apache-2.0).
• Training: LoRA, rank 16, alpha 32, dropout 0.05, learning rate 1e-4 (linear warm-up over 50 steps, then cosine decay over the epoch), batch 8 with gradient accumulation 2, seed 4343. Targets: the attention and Gated-DeltaNet projections of all 32 language layers (152 matrices). The LoRA is merged into the weights and stored in bf16. No other tensor was trained.
• This body is: the checkpoint at 0.75 of the single epoch (step 2,295 of 3,060), chosen from the frozen candidates by the fixed rule (largest mean gain on our two primary dev sets among the pass/fail survivors; 1.82928 vs 1.82879 for the final-epoch body). No average-back arm.
• Loss: cross-entropy on the labelled rows (all of this body's training rows carry dataset labels or computed answers); KL to the starting average's own precomputed probabilities on the regularisation rows (the stored rows carry labels, but the pin KL ignores them). There is NO teacher-distillation term, because this body uses no teacher-probability rows.
• Length limit and loaded dose: rows longer than 2,048 tokens were skipped, not truncated. Of the 34,687 input training rows, 548 were skipped, so 34,139 training rows were actually loaded; of the 14,819 regularisation pins, 2 were skipped, so 14,817 pins were loaded (mean 388 tokens/row). This body is the 0.75-epoch checkpoint, written at step 2,295 of 3,060.
• Relative change of the language tensors from Qwen/Qwen3.5-4B, as Zils computes it: 0.02792, within the Zils 0.15 cap.

Training data (34,687 input rows; 34,139 loaded after the token-length skip — see "Length limit and loaded dose")

Same row count as zils-mini-body-1; the only change is that this body's reasoning/NLI portion is trained on public dataset labels with no teacher probabilities, where the sibling used teacher-distillation rows.

• No-teacher reasoning / NLI / retrieval / language / tools rows (19,583). Public-dataset items carrying their own hard labels (cross-entropy), with no teacher-probability fields of any kind. Drawn by a fixed hash order (not by score) from a pool of 186 public tasksource-style sources, filtered to the families Zils's public exam-authoring code lists as known for the mini recipe (its k2 language/retrieval/tools and k3 reasoning/arts groups), and deduplicated — by exact normalised state, state+question, and 8-gram — against our development sets, the published Zils seed sets v1–v3 (v4 is not in our local inventory and was not checked), and our other training files. Source-text licences as the sources record them: Apache-2.0, CC BY 4.0, MIT, BSD, CC0 and compatible terms. This repository contains model weights only, no training text; share-alike terms (CC BY-SA) apply to redistributing source text, not to these weights.
• Dated outcome questions (8,604). Asked 2026-10-08 to 2026-10-17; every one resolved before 2026-10-20 06:00 UTC (latest 05:59:23); nothing resolved later was used. Rendered in the public Zils Current/Outcomes item format. Wikipedia edit outcomes in 9 language editions 4,952 (CC BY-SA 4.0, attribution via each edit's page history), Wikidata edit outcomes 1,152 (CC0 1.0), arXiv listing categories 2,500 (CC0 1.0). Revert labels re-checked over 72 h.
• Tool questions (4,500). Generated by our own generators with computed answers (tool choice, no-tool, argument scoring, 1,500 each). The generated tool rows reuse public Zils seed instruction templates with new instances; no seed item is used.
• Labelled reasoning (2,000). GSM8K train split (MIT) — the original train rows, not any seed rewrite of them.
• Regularisation (14,819 questions; labels ignored). Public-dataset questions from commercial-use sources (pin set 8365a01d); the target is the starting average's own probabilities (KL-to-start).

Fair-play checks

• The 19,583 new no-teacher rows were deduplicated — exact normalised state, state+question, and 8-gram — against our development sets, the published Zils seed sets v1–v3 (v4 not in our inventory, not checked) and our other training files (the M2 and M3 feeds): 0 matches.
• The other 15,104 rows are deliberately REUSED from our M3 feed (dated outcome questions, generated tool questions, GSM8K). They are not "new" rows, so the no-overlap-versus-M3 claim does not apply to them; they are disclosed here as retained. Those rows passed the same dev/seed dedup when the M3 feed was built (0 matches vs our devs and seeds v1–v3).
• Sources of items in the public Zils exam-authoring code, and the exam's held-out datasets, were excluded; so were the counting-recipe and tool-policy families the mini recipe does not list.
• No teacher-distillation ("relay") rows are used in this body at all.

Our own numbers (not Zils's exam; our own development sets, labelled as ours)

Our own development sets, scored with Zils's open-source grader (mean of 5 secrets). Not the Zils exam. On them, adding a Matilda-start body looks about 1.5 points better than it did in the two published d3-mini rounds, where the real selection did not keep it — read any gain as an upper side, not a forecast.

	mini_combo50 (999)	mini_dev_gold (299)	dev_v2 (3,274)	dev_gold (787)
this body, alone	75.07	79.58	62.86	78.94
{torchcast, d3-mini} (the 11-05 soup)	71.83	76.02	61.38	77.23
{torchcast, d3-mini, this body}	73.66	77.85	62.68	78.42

Our own soup-round check (disclosed in full)

Before this body was prepared for commit we ran Zils's own soup-round logic (greedy +0.5, leave-one-out seat, crown rule) on our development dumps against the field committed for this round. The greedy starts from the single strongest body (astrid-m1, u174). An early run of ours was invalid (it skipped the copy/twin step), so we do not rely on it; the corrected run, on the real round pool (this body on UID 63, our earlier sibling body absent), is complete. In it the greedy selects this body into the soup next to astrid-m1 on one of our two development sets (mini_combo50), but this is not a paid seat — its leave-one-out there is +0.79, under the 0.8 that Zils requires to pay a seat — and on the other set (mini_dev_gold) it is rejected: its trial soup gain there is −0.97, so the greedy does not add it (a rejected body has no leave-one-out). Both decision-round settings agree, and astrid-m1 takes the crown in every cell. So on our development sets this body does not earn a paid seat on either set (paid-seat false in all four cells), but it is borderline — much closer to being a soup member than our earlier sibling body, which added only +0.08 next to astrid-m1. It is entered as a deliberate owner free try; the graded round is the real test. This body sits near astrid-m1 on our twin measure — temperature-fitted Jensen–Shannon divergence 0.0231 bits on mini_combo50 (0.0286 on mini_dev_gold), both above the 0.005 line Zils treats as a twin. It is committed on UID 63 as a deliberate owner free try, replacing our earlier body-1 commitment on that hotkey. On our own dev sets it is a good model on its own (the solo numbers above).

Licence, attribution and modifications

Released under the Apache License 2.0 (see LICENSE). The weights are a modification — a uniform average plus a merged LoRA fine-tune — of the Apache-2.0 works named above; the accompanying NOTICE file states the modification and carries the upstream attributions (Apache-2.0 §4(b)). No weights of any other Bittensor miner's committed body were used.