---
license: apache-2.0
base_model: Qwen/Qwen3.5-0.8B
tags:
- zils
- decision-model
- d3-lite
---

# zils-lite-v2 (v6) — a trained soup body for the Zils `d3-lite` pool (Bittensor subnet 127)

A LoRA fine-tune (rank 64, merged into the weights) of **Qwen/Qwen3.5-0.8B** at revision
`2fc06364715b967f1860aea9cf38778875588b17`, trained from the plain Qwen weights through the frozen
answer layer of the `d3-lite` chassis. The vision tower is the chassis's, unchanged. Language and
vision weights are stored in bf16 in the chassis layout. Not a re-upload of any public model.

Training data, 315,936 rows over 19,746 steps at learning rate 5e-5: the public Zils Relay answer
set (d3-ultra probabilities on bank-style decision tasks), labelled rows from the
`tasksource/tasksource-jev-typed-decisions` dataset that the relay did not use, and real-outcome
decision data (Stack Exchange closes and accepted answers, Wikipedia revert and review outcomes,
Wikidata reverts, arXiv categories, Gerrit merges) with per-family calibrated targets.

Relative change of the language layers from the Qwen parent: 0.048.