zils-lite-body-19: a full fine-tune change, scaled by 1.5, on the current Zils LITE soup

A decision body for the Zils d3-lite soup pool (Bittensor subnet 127), Qwen/Qwen3.5-0.8B family.

This body is below our own pre-set bars (see the read below). We submit it to learn its real score.

It is the current d3-lite soup plus one change of our own:

• Start. The d3-lite soup of round d3-lite-2026-11-10 (soup hash 4d23ec5a54c2ab6d0c16fc493819ef113c789ab8edd1ef8756d41b645d2762a3), which is still the soup with the same two members today, rebuilt by us with the official Zils soup code (soup.build) as the uniform average of its two members, Charlie320/zils-lite-v9a @93b064048a7aeb7cf0d64236136e12c903e67f40 (Apache-2.0) and qhoenix/zils-lite-decision-v9 @083d5b9aa3d2db48155cd6e7f92ac6f9aa057c65 (Apache-2.0). The rebuilt soup's hash equals the published one. Our rebuilt start has Zils body digest 8bebfcbebee0c6fef293680fc44d0f05c2f05a8c014f37eb055602c157cb58dd (weights file sha256 4d456d8c77bb63d04bed623f7f15d75678f9006de17e926e3f82e7c94c9af123). Zils's mining guide allows this ("You may fine-tune the current soup").
• Our change. One full fine-tune of that start on our own data (below); the body is the start plus 1.5 x (the fine-tuned body minus the start).

Credit for the starting weights belongs to the authors of the two soup members and the works they were built on (see NOTICE). Nothing else from any other miner is in this body; it is not a re-upload of any public model. No other miner's outputs were used as labels.

Method

• Training (one run, seed 5062, from the start above). Full fine-tune (no adapters), trainer train_full_gold_v3.py. Learning rate 4e-6, 1 epoch, batch 8 with 2 accumulation steps, on 2,500 labelled rows plus 290 anchor prompts.
  • Labelled rows (2,500): cross-entropy on the row's gold label with label smoothing 0.05, plus a Brier term on choice and yes/no rows or a ranked-probability (RPS) term on score rows (weight 1 each), plus a KL term with weight 0.1 to the start's own answer distribution, computed live from the start weights.
  • Anchor prompts (290): prompts only; their only loss is a KL term with weight 1.0 to the start's own answer distribution, computed live. No labels.
• This body. body = start + 1.5 x (run - start), every tensor (238 tensors differ from the start), built with our weight-difference tool (delta_combo.py v1.4, sha256 0a9b49b1...). No other body merged.

Training data (2,500 rows, file sha256 7e159a477438c85c85cd3197a8712f4e48cba1e9fe8d454803247595abe1e317)

Rows in the exam's 15 "known" families (5 categories x choice / yes-no / score; the Tools families are our dialogue-based versions).

category	choice	yes/no	score	total
reasoning	216	216	215	647
language	216	216	215	647
retrieval	167	167	166	500
tools	153	152	152	457
arts	83	83	83	249
total	835	834	831	2,500

• Most rows were written by AI writers of our team. 2,426 rows (97.0%) were written by two Claude sessions of our team (1,458 and 968 rows) with the public Zils item-authoring recipe (scripts/author_exam.py @6809491), and 5 rows (0.2%) by a local model (Phi-4) following that recipe's method (their records name the style example in seed_source / seed_row_id and carry no recipe field). Their label is the writer's own answer; there is no human label. The Claude-written rows were used only after a second model answered them blind and matched the writer's answer (rows it flagged as unclear were dropped) and after our row guard and audit; the 5 Phi-4 rows (stock re-lay rows) were checked by our own reads instead.
• Style examples. Each Claude-written row was written with one style or topic example: a public TRAIN row (2,399 rows) or an item from our own code generators (27 rows). The rows are new writing, not copies, but we did not check every written row against its own example; what we did check is listed under "Copy checks" below.
• Rows from datasets (69, 2.8%). Their labels are the dataset's own labels or are derived by code from the dataset record (table below).
• Builder. our feed builder (sha256 e7f56b45), from ledger cb388464; difficulty band and hard-row exceptions as recorded in the receipt (exception file 1256f0f8).

source	rows	label	original licence	original release
written by our Claude sessions (public recipe; one public TRAIN row or one of our own code-generator items as style example)	2,426	the writer's own answer; blind-answer match + audit	our own writing	-
written by a local model (Phi-4; the recipe's method, answer steered for label balance; style example: GoEmotions TRAIN 3, TweetEval emotion TRAIN 2)	5	the writer's own answer; checked by our own reads	our own writing	-
SQuAD 2.0 (TRAIN split): answerable-part count	30	dataset gold: answerability of each part, counted by code	CC BY-SA 4.0	https://rajpurkar.github.io/SQuAD-explorer/
Schema-Guided Dialogue (DSTC8, TRAIN split), re-formed to the exam's layout by code	24	code: required slots of the active intent in the human-annotated dialogue state	CC BY-SA 4.0	https://github.com/google-research-datasets/dstc8-schema-guided-dialogue
PIQA (TRAIN split)	13	human gold label: the shown method is the correct or the wrong solution	AFL-3.0	https://yonatanbisk.com/piqa/
SNLI (TRAIN split)	2	human gold label	CC BY-SA 4.0	https://nlp.stanford.edu/projects/snli/

Style examples of the 2,426 Claude-written rows (shown to the writer as a style or topic example; never a label): Schema-Guided Dialogue 433, QNLI 274, RTE 216, Social IQa 216, GSM8K 215, MultiNLI 214, PIQA 203, 20 Newsgroups 167, COCO captions 116, STS Benchmark 99, TweetEval 88 (emotion 78, from SemEval-2018 Task 1; sentiment 10, from SemEval-2017 Task 4), News Headlines for Sarcasm Detection 83, SST-5 73, our own code generators 27, SQuAD 2.0 2. For 66 written rows the example was one specific TRAIN row (20 Newsgroups 23, PIQA 22, Social IQa 21), used as topic and style example only.

Program-built and team-written shares.

kind	rows	share of 2,500
Written by our AI writers (Claude 2,426, Phi-4 5)	2,431	97.2%
Labels computed by code from dataset records (SQuAD 2.0 30, SGD 24)	54	2.2%
Human dataset labels (PIQA 13, SNLI 2)	15	0.6%

Licence notes. 56 rows carry dataset text under CC BY-SA 4.0 (SQuAD 2.0 30, SGD 24, SNLI 2); the rows are not distributed with this body. For the style-example datasets below (shown to the writer; none of them has a row of its own in the training data):

• GSM8K: MIT. COCO captions: the caption annotations are CC BY 4.0 (the images have separate terms; no image was used). News Headlines for Sarcasm Detection: CC BY 4.0 (as its author's Kaggle listing states).
• 20 Newsgroups, RTE: no open licence is named by the original release; used under the original authors' terms.
• SST-5, STS Benchmark: we could not read the terms of the original release (its pages were not reachable); its original terms remain unverified.
• TweetEval emotion (80 rows: 78 Claude, 2 Phi-4): written by an AI writer with a SemEval-2018 Task 1 (via TweetEval) tweet shown as a style example only; no row shares a run of 5 words with its own example tweet, and no training row or anchor prompt shares a run of 8 words with any TweetEval emotion TRAIN tweet (see Copy checks); that dataset's terms limit it to research use.
• TweetEval sentiment (10 Claude rows): written by an AI writer with a SemEval-2017 Task 4 (via TweetEval) tweet shown as a style example only; no row shares a run of 5 words with its own example tweet; that task's terms ask users to follow the Twitter terms of service.
• GoEmotions (3 Phi-4 rows): CC BY 4.0.

Copy checks

• Tweet-seeded rows (90: TweetEval emotion 80, TweetEval sentiment 10). Each was compared with its own example tweet (the 3,257-tweet TweetEval emotion TRAIN file, sha256 448fb3d7..., or, for the 10 sentiment rows, the seed text shown to the writer): no shared run of 8 words, nor of 5 words; the longest shared character run is 15 characters (" bottled water ", the spaces included). Against all 3,257 TweetEval emotion TRAIN tweets: no shared run of 8 words in any of the 2,500 training rows or the 290 anchor prompts.
• Test content. Our exam-split guard and test-overlap rule removed rows that share text with test content. This is a filter against test overlap, not a proof that no source text was copied.
• The other written rows were not checked against their own style example.

Anchor set

290 prompts (file sha256 27de0417dcc9b6f09e4a0a2bfc8deab81ce6a4530b8c7de7e6104ec6aba0c5d3), prompts only. Their targets are the start's own answers, computed live; no labels were used. None of them is in the training rows.

Filters on every row. No exam split files, no public Zils seed items, no Glaive data, no BIG-bench / MMLU; our exam-split guard and test-overlap rule removed rows that share text with test content. Our 149 hold-out rows share no id, no identical state text and no identical state + question with the training rows, and none of the 294 rows of our sealed hold-out list is a training row (by id).

Honest limits of this data. 97.2% of the rows were written by AI writers and carry the writer's own answer, not a human label: the 2,426 Claude-written rows were checked by a second model's blind answer and a sampled audit, the 5 Phi-4 rows by our own reads; none by people. The difficulty band was measured on our rebuilt start.

Build record and our own read (not an official score)

• This body. model.safetensors sha256 60b0ec78aac2b86a6daa3f795bcc398207129e32e95366889abfec69f1857e45. Run weights (seed 5062) sha256 4f374900e73cb7825135bc0e76d524057bf9069a047d956102fb217f4dce31fc.
• Copy screen (official sketch method, current soup kept as the soup): cosine to the soup 0.99791 (line 0.999); share explained by any of our other bodies or the field at most 0.005 (line 0.30). Behaviour distance to the soup and to our other d3-lite bodies at least 0.013 bits (our line 0.01).
• Our own reads (not the Zils exam), against the soup (our rebuilt S10):
  • Exam-style read (C3, 10 registered splits): mean -0.70 points, 3 of 10 splits above 0.
  • Broad dev read (3,279 items): accuracy change -1.37 points (standard error 0.51).
  • Hold-out read (mean change in the probability of the gold answer, our sealed hold-out rows): +0.0515 (passes our hold-out line).
  • Per-family check: fails (some families fall more than 3 points; the largest drop is arts yes/no).
• Our commit bar, and this body. Our pre-set bar is a 10-split mean gain of at least +2.00 points with at least 8 of 10 splits at +2.00 or more. This body is below our own pre-set bars; we submit it to learn its real score.

Licence

Apache-2.0 (LICENSE). See NOTICE for the upstream works and their notices. Nothing here is legal advice; check the original datasets' terms for your use.