zils-mini-hm2

A Zils d3-mini soup-pool body (Bittensor subnet 127): the pinned d3-mini chassis layout (Qwen3.5-4B language and vision tensors, BF16). It succeeds Zayaan/zils-mini-hm1 and adds skill on Zils Outcomes tasks: real-world events whose outcome is checked after the fact. It is evaluated on the pool's pinned chassis, which supplies the answer layer, prompt and runtime.

Method

• Start: zils-mini-hm1 at 584a6787, itself the uniform average of four Apache-2.0 house models (vllm-sr/d3-mini, perplexity-ai/pplx-decider-flash-v1.2, Maincode/Matilda-Decision-v1-Mini, PelaAI/KnowLine-4B-Gen5) fine-tuned with LoRA on verified answers.
• Fine-tune: one epoch of LoRA (rank 32, alpha 64, learning rate 1e-4, seed 11) on the 152 language projections of the Zils d3 runtime with its fixed answer readout (zils.teacher train --start, Zils at 00a72ec), merged into the weights in BF16.
• This body: the uniform average of hm1 and that fine-tune, built with the Zils soup builder (zils.soup.build), so it carries half of the fine-tune's change. Locally it kept hm1's skills on the older items better than the full fine-tune did (see below).
• Targets: the checked outcome of each item, one-hot.

Training data

6,220 rows, 200 of them held out:

• 4,720 Zils Outcomes-format items built with zils.current (Zils at 00a72ec) from public records dated before the window the pool's exams draw from, label-balanced per stream:
  • 956 Gerrit change outcomes (OpenDev, Wikimedia, AOSP), resolved 18–25 Oct 2026;
  • 810 Wikipedia Articles for Creation reviews, 12–25 Oct 2026;
  • 2,800 Wikipedia edits (nine languages) and 154 Wikidata edits from 11–25 Oct 2026, labelled by whether they were reverted.
• 1,500 of hm1's own training answers, so its earlier skills stay.

The export's overlap check found no shared passages with the local evaluation items below, the earlier training set or the public seed exams. 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)

Scored with the validators' calibrated Brier skill (soup_grade.evaluate, mean of five calibration splits):

• Older-style items: 1,191 held-out decision items written like hm1's training data, including the 299 public seed items of zils-ai/base-exam-seed-v1 to -v4.
• Outcomes items: 3,557 items from two validator-size draws (zils.soup_current) of an Outcomes bank built on 9 Nov 2026 from events resolved 24 Oct to 8 Nov.

	hm1	This body
Older-style items (1,191)	75.04	75.15
of which public seeds (299)	86.04	85.41
Outcomes items (3,557)	11.32	15.93

Validator scores are on a private exam and will differ.

Licence

Apache-2.0, as hm1, the four house models and Qwen/Qwen3.5-4B. Training text came from Wikipedia and Wikidata (CC BY-SA 4.0 and CC0) and public Gerrit servers; credit to their contributors.