zils-lite-astrid-g1

A Zils d3-lite (z3-lite 0.8B) soup body: a full fine-tune of vllm-sr/d3-lite (Apache-2.0; itself a fine-tune of Qwen/Qwen3.5-0.8B), trained by us. It is committed to Bittensor subnet 127 (Zils) by the hotkey named in zils-owner.json.

Training

• Data. 5,592 decision items in the Zils exam format. We wrote them with the public authoring recipe in ooo-hq/zils scripts/author_exam.py (author model plus a blind verifier). The style sources were the train splits of public datasets with commercial-use licences:
  • cosmos_qa (CC-BY-4.0), commonsense_qa (MIT), ai2_arc (CC-BY-SA-4.0)
  • gsm8k (MIT), SVAMP (MIT), multi_nli (OANC / CC-BY-3.0 / CC-BY-SA-3.0 / MIT)
  • scinli (Apache-2.0), snli (CC-BY-SA-4.0), dbpedia_14 (CC-BY-SA-3.0), squad_v2 (CC-BY-SA-4.0)
  • boolq (CC-BY-SA-3.0), drop (CC-BY-SA-4.0), hermes-function-calling-v1 (Apache-2.0), go_emotions (Apache-2.0)
  • Sarcasm News Headlines (CC-BY-4.0), amazon_polarity (Apache-2.0), math_qa (Apache-2.0), ropes (CC-BY-4.0), qasc (CC-BY-4.0)
• Exclusions. We did not use the held-out val/test splits, the glaive tools source or the public exam seeds. Items were deduplicated (exact and near-duplicate) against the seeds and the held-out splits. 342 items whose source text also appears in a held-out split were dropped.
• Recipe. Full fine-tune with a soft cross-entropy plus Brier loss and a KL term (0.1) to the base, at lr 5e-6 for 2 epochs.

Evaluation and caveat

• Seeds. On the public exam seeds (zils-ai/base-exam-seed-v1..v4, lite-relevant items) the solo score is 47.3, against 44.45 for d3-lite.
• Outcomes. On a held-out window of the public Zils Outcomes streams (python -m zils.current) this body scores below d3-lite. It is tuned to the authored decision-family exam, not to Outcomes items.

Format

model-00001-of-00001.safetensors plus model.safetensors.index.json are the weights in the d3-lite chassis layout. zils-owner.json is the committing hotkey, as docs/mining.md requires.