zils-lite-body-5

Repository: tensor-tailor/zils-lite-body-5.

A decision body for the Zils d3-lite soup pool (Bittensor subnet 127). It is in the d3-lite chassis layout (Qwen3.5-0.8B language and vision tensors, bf16). It is a fine-tune of the d3-lite pool's 5 November 2026 soup, as 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 in this repository are modified. They are derived from the uniform average of thermopylae-777/zils-lite-body-2 @618946d (apache-2.0), dasLOL/zils-lite-skills-v5 @2c99205 (apache-2.0), then fine-tuned as described below.

Method

• Starting weights: the uniform fp32 average of the bodies above, built with Zils's own soup.build on the vllm-sr/d3-lite chassis at temperature 1.0, exactly as a round builds a soup. These bodies were the 5 November 2026 soup.
• Training: LoRA, rank 16, alpha 32, on the language layers 16-23 only (the top 8 of 24). Learning rate 1e-4, a linear warm-up over the first 50 steps times a cosine decay over the epoch; batch 16; seed 1111; rows longer than 1,024 tokens skipped (training: 151,780 rows used, 2,508 skipped; regularisation: 15,565 used, 293 skipped). The LoRA is merged into the weights and stored in bf16. Every other tensor is identical to the start.
• Loss: cross-entropy on the labelled rows; forward KL to the teacher's option probabilities on the distillation rows; KL to the start's own precomputed option probabilities on the regularisation rows (labels in the stored rows are ignored there).
• Stage 1 ran the single epoch to its final weights (10,460 steps). This is a seat attempt, submitted on the owner's order without our gate checks; our earlier run of the same recipe did not meet our pre-registered crown rule (which rejected every arm of this recipe's first run) or its held-out v4 line. These weights are a re-run of the same frozen recipe (same data, seed and settings); the first run's weights (digest 933ce3871fe79fad3817d7dd35ec77729e744a5518a14a2bab5d5c8a1054ea1e) were lost before upload, and this re-run's weights (digest d2e11c317a4f53961f8922773bbb445f50d4dffc8125fa24b43d041ae1de265d) differ from them.
• Stage 2, which gives these weights: starting from the stage-1 weights, one more LoRA epoch with the same settings (seed 1133, 653 steps, final weights) on 9,600 rows given: 2,399 questions the stage-1 weights answered wrongly in a fresh pool of 21,600 newly generated questions (the same generators with new seeds; no scenario equal to a training scenario, and none sharing an 8-word run with any of our test sets), 1,238 correctly answered rows of the same question types (to keep each type's answers balanced), and 5,963 rows replayed from the first-stage data. Of those, 9,510 were used and 90 skipped by the 1,024-token limit; regularisation: 938 used, 22 skipped. The rows were picked by a rule fixed before any result existed. This second stage was the one pre-declared follow-up. Only final checkpoints were eligible for shipment; no checkpoint was chosen by score. The owner's decision chose between the stage-1 and stage-2 weights, and it chose these stage-2 weights.
• Relative change of the language layers from Qwen/Qwen3.5-0.8B: 0.048 (Zils cap 0.15).

Training data (154,288 rows before the length limit)

The first 80,303 rows come from the mix of our earlier body tensor-tailor/zils-lite-body-4 (81,422 rows there; 1,119 removed here, see Filters and overlap):

• Distillation (17,324 questions). The targets are the option probabilities of vllm-sr/d3-ultra (Apache-2.0) from zils-ai/zils-relay-answers (Apache-2.0). The question text was rebuilt locally from the public sources that card pins: 14,783 via tasksource, 2,259 from Hugging Face datasets and 282 BFCL tool-selection rows.
  • Largest sources: sem_eval_2010_task_8 2,512; multilingual NLI 1,784; twitter-financial-news-sentiment 1,314; babi_nli 1,206; ARC-Easy 637; HellaSwag 635; WinoGrande 630; ARC-Challenge 357; HelpSteer (three versions) 855.
  • Licences of the source texts, as the relay card records them (rows): MIT 3,366; CC0 3,347; CC BY 4.0 3,079; Apache-2.0 2,768; CC BY-SA 4.0 2,152; BSD 1,270; AFL-3.0 623; GPL 303; CC BY-SA 3.0 190; ODC-By 65; ODbL 6; other 155. Splits are as the relay card records them; we did not re-check them for this card.
• Generated questions with computed answers (19,281), from our own generators: tool choice, no-tool and argument scoring, tool-call policy, coverage, counting, word forms, rule-book scoring and syllables.
• Syllables (11,737), our own generator over the CMU Pronouncing Dictionary (BSD-2-Clause-style licence).
• Word problems (12,000), our own generator, own work.
• Intent scope (6,244), our own generator, own work.
• Procedural decisions (13,717) from tasksource/tasksource-jev-typed-decisions @8df2e5c9 (Apache-2.0), train split, with the dataset's own labels.

The other 73,985 rows are new questions with computed answers. They come from our own generators, are our own work, and were checked by an independent re-labeller:

• Calculation (25,200): totals and counts over the same facts written as prose, bulleted lists, JSON inventories, tables and mixed text.
• Inference (9,985): order chains of 5-7 people with distractor facts; category chains of 3-4 steps with decoy rules; location hops 3-5 deep; chained if-then rules.
• Syllables (14,800): words from the CMU Pronouncing Dictionary train split (BSD-2-Clause-style licence). Excluded are the dictionary's test and validation words, every word in any state of the public Zils seed sets v1-v3 and of our dev, held-out and fresh-instance sets, and the syllable words of our earlier training rows.
• Tool-use decisions (16,000): whether the assistant should call a listed tool, ask for a missing value, say no tool can do it, or answer directly, over synthetic tool lists. The four categories follow the definitions in the NVIDIA When2Call README (Apache-2.0). No When2Call data is used.
• Rule-book scoring (8,000): the level a short post earns under a stated word rulebook.

Regularisation (15,858 questions before the length limit): public-dataset questions whose target is the start's own probabilities.

Share-alike and copyleft terms apply to redistributing source text. This repository contains model weights only, no training text.

Filters and overlap

• The first rows (81,422 in tensor-tailor/zils-lite-body-4) were checked when they were built for that body, as its card states. They were checked against the published Zils seed sets v1, v2 and v3 (v4 was not checked) and our dev and held-out sets: 0 exact matches on the normalised scenario + question, and 0 scenario matches with our held-out sets. Here, a stricter check removed 1,119 of them (below); the other 80,303 are unchanged.
• The 73,985 new rows were checked against the published Zils seed sets v1, v2 and v3 (v4 was not checked) and our dev, held-out and fresh-instance test sets, including our held-out v4 set. The check covered an equal normalised scenario and any shared run of 8 words: 0 rows matched. 15 generated rows sharing one generic 8-word template run with 2 held-out items were removed before training.
• Shared 8-word runs with our test sets, the whole mix: a row was removed if it shared a non-template run of 8 words (a run found in fewer than 5 rows of that test set, and not also part of a shared template run) with a test item and had the same answer label. 1,119 rows removed before training: 1,102 by the overlap rule and 17 by a since-fixed checker defect (one-word lemma states); no test item was copied. All 1,119 are rows of the first part (1,004 coverage scoring, 55 tool scoring, 41 counting, 17 word forms, 2 intent scope). The other rows that share only format templates with our test sets (exam boilerplate our generators use, and one-word states with different questions, apart from the 17 removed above) were kept.
• Caveat: because those format templates are shared, this body's dev-set gains in the tools, retrieval, arts and reasoning categories are partly format training. Our held-out v4 set uses different templates and is the honest read; this exact body was not measured on it before submission (it was submitted on the owner's order without our gate checks). No new row equals a row of the first 81,422 or another new row.
• Not used: 20 Newsgroups, Sarcasm Headlines, PIQA, SVAMP, MMLU and VitaminC (sources or held-out sets of the public Zils exam-authoring code), and no real-event (dated outcome) rows at all.
• Some of the earlier generated rows reuse public Zils seed instruction templates with new instances. The 73,985 new rows use our own instruction text. No seed item is used.
• Source lineage (disclosed): 642 of the 154,288 rows come from public datasets that our team later classed as derived from, or seeded by, datasets the Zils exam draws on. Zils's rules do not ban them; this is our own stricter rule, adopted after this mix was frozen. By source: winodict 76, PromptShield 74, winowhy 65, DynaSent R2 65, DynaSent R1 64, resnli 61, multilingual-sentiments 61, dgen 54, equate 45, syntactic-augmentation-nli 40, naturallogic 37. No exam item was used.

Inherited through the starting weights

The starting bodies were trained by others and by us earlier; their own cards describe their data.

Inherited-data bullets for the card, one per likely 11-06 soup member (worker2, 11-05 ~22:50Z)

At T, copy the bullets of the ACTUAL members into inherited.md, then run fill_card.py --inherited inherited.md.

• Each bullet paraphrases that member's own public card at the pinned revision, read anonymously on 11-05. Recheck the card if the revision changed.

• B1's and dasLOL's bullets are taken from our H1 card. dasLOL's had a third sentence ("We did not choose or check that data; see its card."), which is not kept.

• thermopylae-777/zils-lite-body-2 (B1, ours, @618946d): trained from the uniform average of the three d3-lite house bodies (vllm-sr/d3-lite, Maincode/Matilda-Decision-v1-Edge, zils-ai/jiwo-0.8b-full) with relay distillation and generated questions; see its card.

• dasLOL/zils-lite-skills-v5 (@2c99205): its card states six full-parameter fine-tunes from the same house average, averaged, on generated skills and public train splits including some (20 Newsgroups, Sarcasm Headlines) that we do not use ourselves; see its card.

• Charlie320/zils-lite-k1c (@2ae2816): its card states a full-parameter fine-tune from the uniform average of the three house bodies, on project-generated computed questions and public train splits under their own licences; see its card.

• Charlie320/zils-lite-k3b (@71d4ffe): its card states a top-layers fine-tune of the 5 November soup (B1 and dasLOL) with soft targets, computed questions and regularisation to the start, on project-generated questions and public train splits; 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.

• tensor-tailor/zils-lite-body-3 (B3v2, ours, @9b943bd): a merged-LoRA fine-tune of our unpublished body B2, which starts from the uniform average of the three house bodies; its card lists the training data; see its card.

• qhoenix/zils-lite-decision (@9e88637): its card states a LoRA fine-tune of the average of vllm-sr/d3-lite, Maincode/Matilda-Decision-v1-Edge and cribl-ai/cribl-decision-2-0.8B, on train splits of public classification, NLI and QA datasets and resolved public events, averaged 50/50 with that base; see its card. (Date-rule flag, see below.)

• bestdev57/zils-lite-body-1 (@9e2e2ae): its card states a LoRA fine-tune of the d3-lite chassis on resolved real-world outcomes (arXiv categories, Wikipedia reverts, Articles for Creation and deletion outcomes, Gerrit, Stack Exchange) collected before its evaluation window; see its card. (Date-rule flag, see below.)

• BlessJ/zils-lite-body-1 (@2564583): card NOT readable anonymously on 11-05 (the site answered "Invalid username or password"), so the licence is unknown. Per H2_RUNBOOK step 1, a closed or unknown licence stops the run.

Date-rule flag (for the manager and owner, before T). Two candidates state training on resolved real events: qhoenix-169 and bestdev57-182. Our order 6 date rule covers OUR rows: no real-event row resolved on or after 2026-10-20T06:00Z, and our own data has none. If either body is a member of the published 11-06 soup, H1v2 would inherit its weights through the start (e). Whether that is allowed is the owner's call, not mine. The card would state it in the bullet above either way.

Our own measurements (not the Zils exam)

We did not measure this exact body before submission: it was submitted on the owner's order without our gate checks (our earlier run of the same recipe failed our pre-registered crown rule and its held-out v4 line).

Licence

All upstream weights (Qwen/Qwen3.5-0.8B, vllm-sr/d3-lite and the house bodies inside the starting bodies) and the starting bodies themselves (thermopylae-777/zils-lite-body-2: apache-2.0, dasLOL/zils-lite-skills-v5: 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.