zils-lite-body-d2

A fine-tune 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.

Lineage

Training started from an equal-weight average of three Apache-2.0 house bodies, each derived from Qwen/Qwen3.5-0.8B:

No other participant's body was used as a training start or as a data source. Published bodies thermopylae-777/zils-lite-body-2 and dasLOL/zils-lite-skills-v5 were used only as evaluation baselines. The general recipe of a LoRA from a house average with pin, teacher and generated rows follows the one described on the model card of thermopylae-777/zils-lite-body-2.

Training

python -m zils.teacher train --lora 64 --lr 1e-4 --seed 8 --start <the average above> from the Zils repository at 17551565cc6a7636b452765d205526a5f479aa7d: one pass over 43,260 rows, then merged into the weights. Training took about 2.1 hours on one RTX A6000.

• 5,000 pin rows: the starting average's own answers on sampled questions. The first 1,000 are the held-out KL set (final KL 0.109).
• 20,000 rows from zils-ai/zils-relay-answers (Apache-2.0 probabilities; the rebuilt question text stays under its sources' licences).
• 15,000 practice questions generated by the submitter from its own word lists and templates, with answers computed by code. The generators were designed after reading the question formats of the public base-exam-seed families. No seed item text is in the training set (exact-text overlap checked: 0).
• 3,260 rows labelled by public outcomes, built with python -m zils.current from arXiv metadata, public Gerrit servers, Wikipedia Articles for Creation and Stack Exchange. 2,321 are arXiv rows; the other 313 are repeated three times. 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, and no attempt was made to find or infer them.

Self-reported evaluation

Measured by the submitter on a local approximation of the exam, not by a validator. Scores use the protocol's calibrated Brier scoring with category and block weights, renormalized over the available cells.

Bootstrap SD of the gain from adding this model to the current soup: 1.22, 0.62 and 0.60 points.

Limitations

• These are not validator scores. On the public board at the same time, the approximation ranks the current soup below thermopylae-777/zils-lite-body-2 while the board ranks it above, so these points do not convert to board points. It also shows dasLOL/zils-lite-skills-v5 lowering a soup, unlike the board.
• The seed-based columns may be optimistic, because the generators were designed after reading seed formats.
• The gain from adding this model on seeds alone (+1.65) does not clear the greedy margin plus one SD.
• Alone, this model is 2.2 points below the current soup on seeds. Arts drops from 49.6 to 34.2 against the current soup, while tools rises from 38.9 to 50.5.
• Seeds cover only the known block. The new block is represented by two Outcomes families. Wikipedia revert streams are not covered.