Zils lite decision body
Fine-tune of the uniform average of vllm-sr/d3-lite at 03560c26f5199c3c04129a2ccb38960c913635e0, Maincode/Matilda-Decision-v1-Edge at 90c2cee9c6f607f40a4802a95939857d7eb6dfeb, and zils-ai/jiwo-0.8b-full at adce2ce92cd2130fd3ed451cf5f9642e3dfe4a57.
The language weights were trained with a rank-16 LoRA on permissive rows of zils-ai/zils-relay-answers (Apache-2.0), then on a mix of those rows and Matilda's distributions over train-split commonsense questions. Share-alike and non-commercial sources were left out.
CommonsenseQA train is MIT. QASC train is CC BY 4.0. HellaSwag train is MIT. The BIG-bench tasks packaged by tasksource/bigbench are Apache-2.0. Vision weights and the answer layer are the d3-lite chassis.
- Downloads last month
- 31
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support