zils-mini-hm1

A Zils d3-mini soup-pool body (Bittensor subnet 127): the pinned d3-mini chassis layout (Qwen3.5-4B language and vision tensors, BF16). Its start is the uniform average of four of the pool's house models, fine-tuned with LoRA on verified answers. It is evaluated on the pool's pinned chassis, which supplies the answer layer, prompt and runtime.

Method

• Start: the uniform average of four Apache-2.0 house models at pinned revisions, built on the pinned chassis with the Zils soup builder (zils.soup.build), the way validators average soup members:
  • vllm-sr/d3-mini at df99bf4c
  • perplexity-ai/pplx-decider-flash-v1.2 at 9ff7cf98
  • Maincode/Matilda-Decision-v1-Mini at 39479a8a
  • PelaAI/KnowLine-4B-Gen5 at a8e0cd66
• Training: one epoch of LoRA (rank 32, alpha 64, learning rate 1e-4) on the 152 language projections of the Zils d3 runtime with its fixed answer readout (zils.teacher train --start), merged into the weights in BF16. No other tensor was trained.
• Targets: the verified answer of each item.

Training data

4,992 exam-format decision items written by Claude Opus 5.5 with the Zils exam authoring recipe (scripts/author_exam.py: one public row as a style example, an independent blind answer, kept only on agreement), using public training splits only:

• 2,974 items for the 15 current known families: Social IQa, PIQA, SVAMP and GSM8K, GLUE MNLI, RTE, QNLI and STS-B, Crisscrossed Captions (validation), 20 Newsgroups, Glaive function calling v2, dair-ai/emotion, News Headlines Dataset for Sarcasm Detection, and TweetEval sentiment;
• 2,018 items across 99 BIG-bench training tasks (tasksource/bigbench) and Glaive function calling.

No validator exam items, no public seed items and no customer data were used. This repository holds model weights only, no training text; check each source's terms before reusing it.

Local evaluation (not the validators' exam)

1,191 held-out items, scored with the validators' calibrated Brier skill (soup_grade.evaluate, mean of five calibration splits): 892 written the same way from rows never used in training, including 15 BIG-bench tasks kept out of training, plus the 299 current-family items of zils-ai/base-exam-seed-v1 to -v4.

	Start (four-model average)	This body
All 1,191 items	68.70	75.10

On the same items three of the four house models alone score 64.20 (d3-mini), 60.17 (Matilda-Decision-v1-Mini) and 58.44 (KnowLine-4B-Gen5).

Validator scores are on a private exam and will differ.

Licence

Apache-2.0, as the four house models and Qwen/Qwen3.5-4B. Credit to their authors; please cite BIG-bench (Srivastava et al., 2022) and tasksource (Sileo, 2024) for the training sources.