d3-lite body a3
A Zils d3-lite body (Qwen3.5-0.8B decision model). Trained from the published d3-lite soup with a rank-16 LoRA (one epoch, AdamW lr 2e-4, linear schedule) on public typed-decision rows plus generated multi-step arithmetic counting items, with a KL anchor to the starting soup's answers; the LoRA is merged into the weights.
- Downloads last month
- 8
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support