zils-lite-mixed-v1
A fine-tune of vllm-sr/d3-lite at revision
03560c26f5199c3c04129a2ccb38960c913635e0 (itself derived from Qwen/Qwen3.5-0.8B), submitted as a
soup body to the d3-lite pool of Zils (Bittensor subnet 127). It holds weights only; it runs on the
pool's pinned chassis and answer layer.
Training
python -m zils.teacher train --lora 64 --seed 2 from the Zils repository at
17551565cc6a7636b452765d205526a5f479aa7d: one pass over 39,617 rows, then merged into the weights.
- 37,000 rows from
zils-ai/zils-relay-answers(Apache-2.0 probabilities; the rebuilt question text stays under its sources' licences). - 2,617 rows labelled by public outcomes, built with
python -m zils.currentfrom arXiv metadata, public Gerrit servers, Wikipedia Articles for Creation and Stack Exchange, asked in late October 2026. Those sources carry their own terms (for example CC BY-SA for Wikipedia and Stack Exchange text). No question text is redistributed here.
No customer data and no private exam items were used.
Self-reported evaluation
Measured by the submitter, not by a validator. Paired against stock vllm-sr/d3-lite, with bootstrap
95% intervals, on one RTX A6000:
| Check | Stock | This model | Paired gap |
|---|---|---|---|
| Brier skill, six Outcomes families with at least 40 held-out items (2,554 items) | 42.5 | 56.9 | +14.4 [+11.5, +17.5] |
Brier loss, 38 public base-exam-seed families (754 items, unseen in training) |
0.4255 | 0.4263 | -0.0008 [-0.0108, +0.0091] |
| Accuracy, the same seed families | 66.7% | 68.6% | not bootstrapped |
Limitations: four of the six Outcomes families are arXiv category questions, and training used other items from the same streams, so most of the gain is specific to those streams. Wikipedia revert streams were not resolved in time and are not covered. On the unseen seed families the model is level with stock, not better. The private election exam may give a different result.