zils-lite-body-10: class-balanced human-label fine-tune of the current Zils LITE soup

A decision body for the Zils d3-lite soup pool (Bittensor subnet 127), Qwen/Qwen3.5-0.8B family. It is a fine-tune of qhoenix/zils-lite-decision-v9 @083d5b9aa3d2db48155cd6e7f92ac6f9aa057c65 (Apache-2.0; Zils body digest 8b8e3db3c8a62cbde4b5fb034bffb2c871faa1db450f6ddb0103ad120a44cfe5), which is the current d3-lite soup, as Zils's mining guide allows ("You may fine-tune the current soup"). Credit for the starting weights belongs to its author and the works it was built on (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 attention and Gated-DeltaNet projections of every language layer (no MLP), merged into the weights after training (trainer train_soft_c_gold_v3.py bc191000). Learning rate 2e-5, AdamW with weight decay 1.0 on the LoRA weights, batch 8 x accumulation 2, warm-up 50 steps then cosine, 2 epochs, max 2,048 tokens, seed 4343; loss per row = cross-entropy on the row's gold label + 1.0 x Brier over its options. No KL term to the start, no teacher or reference outputs.
• This body: the save at 75% of training (1.5 of the 2 epochs) of the LoRA fine-tune, merged into the weights (no weight averaging).

Training data (4,262 rows, file sha256 90d702689f46e427d0ede0ef5a71be93186c631a1da1db3c0b56729f0c9fd39a)

A class-balanced set in the exam's 15 "known" families (5 categories x choice / yes-no / score; category shares follow the published exam weights where a cell had enough rows). Every label is the dataset's own label (human annotators, or the dataset's own prover / generator), derived by code from the dataset record, or computed by our own generator; no label was predicted by a model we ran, and no other miner's outputs were used. Version 3 of our class-balanced LITE set (band difficulty rule below; freeze tool balL_freeze.py v1.8 33a629d3). Difficulty rule per cell (to avoid the label-noise tail of a pure hardest-first pick): rows with a human label were taken in a fixed hash order from those on which the starting body (decision-v9) puts probability 0.30-0.90 on the correct answer (rows at or above 0.90 only to fill a short cell, the hardest of them first); code-labelled and proven rows hardest-first; 50% cap per source except where noted in the freeze receipt.

category	choice	yes/no	score	total
reasoning	431	646	431	1508
language	430	428	473	1331
retrieval	334	332	0	666
tools	175	0	0	175
arts	166	166	250	582

source	rows	label	original licence	original release
PIQA (TRAIN split)	646	human gold label: the shown method is the correct or the wrong solution of the goal	AFL-3.0	https://yonatanbisk.com/piqa/
MultiNLI (GLUE, TRAIN split)	430	human gold label (3-way; in yes/no rows entailment = yes, neutral / contradiction = no)	OANC licence; fiction parts CC BY-SA 3.0 / CC BY 3.0	https://cims.nyu.edu/~sbowman/multinli/
Social IQa (TRAIN split)	378	human gold answer	CC BY 4.0	https://allenai.org/data/socialiqa
20 Newsgroups (SetFit mirror, TRAIN split)	334	the newsgroup the post was sent to	no open licence is named by the original release; used under the original authors' terms	http://qwone.com/~jason/20Newsgroups/
STS Benchmark (TRAIN split, captions genre)	323	human 0-5 similarity mean mapped to the nearest of 5 levels; rows near a level edge dropped	no open licence is named by the original release; used under the original authors' terms (mixed sources; the mirror card says 'unknown')	https://huggingface.co/datasets/mteb/stsbenchmark-sts
SST-5 (Stanford Sentiment Treebank, TRAIN split)	250	human 5-class sentiment label	no open licence is named by the original release; used under the original authors' terms	https://nlp.stanford.edu/sentiment/
SNLI (TRAIN split)	215	human gold label	CC BY-SA 4.0	https://nlp.stanford.edu/projects/snli/
RTE (GLUE, TRAIN split)	213	human gold label (entailment / not entailment)	no open licence is named by the original release; used under the original authors' terms (PASCAL RTE challenges)	https://gluebenchmark.com/
QNLI (GLUE, TRAIN split; text from SQuAD v1.1)	167	human gold label (the sentence answers the question)	CC BY-SA 4.0 (SQuAD)	https://rajpurkar.github.io/SQuAD-explorer/
iSarcasmEval task A (TRAIN split)	166	the tweet author's own sarcasm label	MIT	https://github.com/iabufarha/iSarcasmEval
SQuAD 2.0 (TRAIN split, sentence level)	165	derived by code from the human answer spans (the sentence holds the answer or not)	CC BY-SA 4.0	https://rajpurkar.github.io/SQuAD-explorer/
ToolACE @6bda777c	100	read off the record: the real call is the answer; distractors break a required argument (code-checked)	Apache-2.0	https://huggingface.co/datasets/Team-ACE/ToolACE
GoEmotions (simplified, TRAIN split)	83	human label (rows with exactly one of the 6 emotion names)	Apache-2.0 (google-research repository)	https://github.com/google-research/google-research/tree/master/goemotions
TweetEval emotion (SemEval-2018 Task 1, TRAIN split)	83	human emotion label	no open licence is named by the original release; used under the original authors' terms (TweetEval: 'restrictions may apply to individual tasks'; Twitter terms)	https://github.com/cardiffnlp/tweeteval
Hermes function-calling v1 @dae3e1d2 (func-calling file; the authors' Glaive-derived glaive-function-calling-5k file excluded)	75	read off the record: the real call is the answer; distractors break a required argument (code-checked)	Apache-2.0	https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1
RuleTaker (tasksource mirror)	53	the dataset's prover label under the stated closed-world assumption	Apache-2.0 (allenai/ruletaker repository)	https://github.com/allenai/ruletaker
our code generator: reasoning score items (ASDiv-style word problems)	431	computed by the generator from its item spec and re-counted independently	our own code	-
our code generator: language score items	150	computed by the generator from its item spec and re-counted independently	our own code	-

Licence note. 1,203 of the 4,262 rows come from 5 datasets whose original release names no open licence (20 Newsgroups (SetFit mirror, TRAIN split), STS Benchmark (TRAIN split, captions genre), SST-5 (Stanford Sentiment Treebank, TRAIN split), RTE (GLUE, TRAIN split), TweetEval emotion (SemEval-2018 Task 1, TRAIN split)). We used their TRAIN splits for training only, under the original authors' terms; this repository contains model weights only, no dataset text. Check those datasets' terms before you reuse this body for anything they may restrict.

Filters on every row: no exam split files (the validation / test splits the Zils exam-authoring code reads were never opened), no Glaive data, no BIG-bench / MMLU, no real-event rows; our exam-split 8-gram guard and test-overlap rule (both directions) removed every hit; near-duplicates removed; no rows written by an LLM item writer, no house or relay labels, no other miner's outputs as targets. Share-alike terms (CC BY-SA) apply to redistributing source text; this repository holds model weights only.

Checks we ran before commit (not an official score)

• Change from Qwen/Qwen3.5-0.8B (Zils relative change): 0.03556 (cap 0.15). Zils body digest fe50bb3aa8a5c8f5ac769c9bb263d9f01290ae0a66a936c704708de37db7f201.
• Copy screen with the official soup_copies code (our seeds; the validator's seed is private): own share beyond the house bodies, the listed public models, the soup and the release 0.018-0.019, i.e. below (house-mix line 0.05); largest cosine to any of them or to our other committed bodies 0.991 (with the soup body qhoenix/zils-lite-decision-v9) (line 0.999); largest share explained by earlier bodies 0.004 (line 0.30).
• Our own exam-style reads vs the start (same code, same items; not the Zils exam): predicted score +0.15 points overall, within our measurement error (flat); by category reasoning -2.1, language +3.8, retrieval -3.7, tools +3.1, arts +0.2. On the known-family part alone reasoning (-1.77), retrieval (-6.98) and arts (-2.32) fell; the retrieval drop is over-confidence (its accuracy is flat). This 75% save was chosen after the reads; it was not fixed in advance.

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.