zl-mini-o2

A d3-mini soup body for Zils (Bittensor subnet 127).

  • Start: the current official d3-mini soup (round d3-mini-2026-11-12, soup sha256 9a21f188โ€ฆ), rebuilt with zils.soup.build from qhoenix/zils-mini-decision-l13_avg678 @ 091c6eb and joamCriff123/zils-mini-sx3 @ c9432ad.
  • Training: LoRA rank 32, lr 4e-5, one epoch, soft cross-entropy plus Brier loss with a KL anchor (0.3) to the start.
  • Data: about 15,900 exam-format items. About 9,800 are resolved real-world outcomes collected with zils.current from open CC0 / CC-BY-SA sources (Wikipedia and Wikidata edit reverts, AfD and AfC outcomes, arXiv categories), labelled by what actually happened. The rest are program-verified arithmetic items (questions seeded from the GSM8K train split, written by gpt-oss-20b, labels computed by executing code) and earlier items for the other known families written by open-weight models. Items were deduplicated against held-out evaluation data by source row, normalized fingerprints, option order and embedding similarity.

Weights are bf16 safetensors in the pool's layout. License: Apache-2.0, as its parents.

Downloads last month
26
Safetensors
Model size
5B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for jwcwn/zl-mini-body-5

Finetuned
(5)
this model