zils-lite-body-7

Repository: tensor-tailor/zils-lite-body-7.

A decision body for the Zils d3-lite soup pool (Bittensor subnet 127). It uses the d3-lite chassis layout (Qwen3.5-0.8B language and vision tensors, bf16). It is a full fine-tune of the d3-lite pool's 6 November 2026 soup, which Zils's mining guide allows ("You may fine-tune the current soup"). It is not a re-upload of any public model.

Modified weights

The weights here are modified. They derive from the uniform average of 1lock/zils-lite-astrid-g1 @ac979f3 (apache-2.0) and bmars8218/zils-lite-body-2 @60b596b (apache-2.0), fine-tuned as described below. Zils body digest of these weights: a64c0cb733201e8a37b32112d6f164eb1abf56d07eaf994657d37da616af7fcb.

Method

• Starting weights: the uniform fp32 average of the two bodies above, built with Zils's own soup.build on the vllm-sr/d3-lite chassis at temperature 1.0, as a round builds a soup. Zils body digest of the start: df5a2e1be25a3d23494dd7d0e57d645376e7c062288ff30bf6067dd19833ce2c.
• Training: full fine-tune of every language-model tensor, held as an fp32 master under bf16 autocast. Vision tensors are unchanged.
  • AdamW (weight decay 0), learning rate 5e-6, a linear warm-up over 50 steps then cosine decay, gradient clip 1.0.
  • Batch 16, seed 4343, 2 epochs (1,486 steps), rows longer than 1,024 tokens skipped.
  • Checkpoints at 25 / 50 / 75 % and the end. The shipped weights are the final ones (step 1,486).
• Loss on the 8,931 training items: soft cross-entropy to the item's answer, plus a Brier term (weight 1), plus KL (weight 0.1) to the start's own probabilities on that item.
• Regularisation (2,977 rows, 25 % of all rows; 2,949 used, 28 skipped by the length limit): KL (weight 1) to the start's own precomputed probabilities on broad public-dataset questions (the same regularisation set as our earlier bodies). It is intended to preserve the start's ability on unrelated questions; the measured losses below show it did not fully do so.
• Relative change of the language layers from Qwen/Qwen3.5-0.8B: 0.021 (Zils cap 0.15; the start: 0.020).

Training data (8,931 items)

All items are in the format of the public Zils exam. Every item comes from a public dataset's TRAIN split.

• 5,856 items written for this body (5 families):
  • The public Zils exam-authoring code (scripts/author_exam.py @ab3a98ef: its rewrite instructions, question types and schema) was run by the writer model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 (Apache-2.0). Each item rewrites one row of a train split.
  • The writer then answered each item blind: greedy plus 4 samples, at least 3 of 4 agreeing with greedy, and the answer had to equal the source row's own label. Everything else was dropped.
  • Families and counts:
    • language-known-noul-k2: 2,364 (from MultiNLI);
    • reasoning-known-noul-k3: 1,506 (from PIQA);
    • language-new-noul-veridicality: 754 (MegaVeridicality);
    • arts-new-noul-amazon-polarity: 734 (Amazon Polarity);
    • retrieval-new-noul-boolq: 498 (BoolQ).
• 3,075 human-labelled rows from train splits, converted to the exam format; the labels are the datasets' own:
  • language-known-choice-k2: 1,257 (MultiNLI, keys contradiction / neutral / entailment);
  • retrieval-known-noul-k2: 1,430 (QNLI);
  • language-new-choice-copa: 388 (COPA).
• Licences of the original datasets, as their authors state them:
  • MultiNLI: OANC licence; fiction part CC BY-SA 3.0 / CC BY 3.0;
  • QNLI: CC BY-SA 4.0, via SQuAD;
  • PIQA: Academic Free License 3.0;
  • COPA: BSD-2-Clause;
  • MegaVeridicality: MIT;
  • Amazon Polarity: Apache-2.0;
  • BoolQ: CC BY-SA 3.0.
  • The public exam-authoring code reads other splits of some of these datasets. No row of those split files is used.
• Balance: yes/no families 50/50; choice families an equal number per answer key.

Gates (an item that failed any gate was removed; no item was edited)

• Seeds were screened before writing. Two independent reviewers then checked every item file, and again the exact final file.
• Item gates:
  • 8-word overlap with the whole Glaive function-calling file: 0.
  • 8-word overlap with the 81 sources the public exam-authoring code reads: 0.
  • Rule-B against the exam's dev and held-out material.
  • Seed keys and 8-word runs against every test set of ours, including older ones; non-template overlap removed.
  • Option counts as in the exam's dev set.
  • No two options with the same text.
  • Near-duplicates removed.
  • One item per source row.
  • Label balance.
• Manual audit: 20 items per family were drawn by a fixed salted hash from the final rows and read by a reviewer.
  • A family with 2 or more confirmed defects was removed whole. A single defect, or an unresolved doubt, removed that item.
  • Every family that met this rule was removed (most of the written families and four converted ones), mostly for answer keys the passage did not support.
• Licence: only datasets whose original licence is open were used. The five datasets held back for our unseen-domain tests (LogiQA, ReClor, WiC, WikiQA, SST-5) are not used.
• Not used: Glaive, BIG-bench, MMLU, WSC-derived sets; no real-event (dated outcome) rows.

Inherited through the starting weights

• 1lock/zils-lite-astrid-g1 (@ac979f3): its card states a full fine-tune on exam-style items written with the public Zils exam-authoring code from training-split rows and kept after a blind check; see its card.
• bmars8218/zils-lite-body-2 (@60b596b): its card states a LoRA (rank 16) fine-tune of the 5 November soup with the official trainer, on relay soft answers (zils-ai/zils-relay-answers), generated rows with computed labels following the public exam seed recipes, and the start's own probabilities as a regulariser; see its card.

Selection and our own measurements (not the Zils exam)

We had pre-registered a forgetting gate and a generalisation gate for the four saves, with an ordered stop rule. They were measured and FAILED for every save, and were not enforced: the owner decided to ship this body anyway. Only the checks Zils requires were kept as shipping conditions:

• the twin distance to the pool bodies;
• the copy check;
• the package identity.

The save was picked by our prediction score on our pool tests (best PRED, ties to the earlier save).

This body scores lower than its start on our dev set, our prediction score, the guard families, the unseen-domain test and the new-domain sets. Its held-out v4 top-choice accuracy is higher: 0.531 against 0.511 for the start. All four saves failed the forgetting gate we had pre-registered; we ship the final save only on the owner's decision.

• Dev score: 45.40 / 48.87 / 45.94 / 47.27 at 25 / 50 / 75 / 100 % of training, against 59.19 for the start.
• Guard families: tools yes/no -0.16 to -0.32, tools choice -0.19 to -0.23, retrieval score -0.19 to -0.26.
• Unseen-domain test: -3.4 to -6.8 points.
• New-domain sets: -2.1 to -8.8 points.

Our likely explanation: the training mix has only 8 families (~9k rows), so the full fine-tune overwrote the skills it did not see.

	Reasoning	Language	Retrieval	Tools	Arts	Dev score (total)	K	N	PRED
this body (final save)	33.64	51.42	67.89	38.45	46.59	47.27	51.33	73.56	60.22
the start (6 November soup)	38.41	70.62	77.05	56.63	52.32	59.19	58.77	77.40	66.22

PRED per save: 58.68 / 57.32 / 59.57 / 60.22 (start 66.22). All numbers are from our own test sets (not the Zils exam); the start was evaluated on the same machine.

Licence

The upstream weights (Qwen/Qwen3.5-0.8B, vllm-sr/d3-lite and the bodies inside the starting soup) and the starting bodies themselves (1lock/zils-lite-astrid-g1: apache-2.0; bmars8218/zils-lite-body-2: apache-2.0) are released under the licences their cards state. Credit to their authors. This repository is licensed under Apache 2.0; see LICENSE and NOTICE. Nothing here is legal advice; please check the original datasets' terms for your use.