starlight trinity nano
Very early WIP !!
extremely minimal 300 steps / 60 data. this is a training test, not a functional finetune (yet!)
proved training will work if i scale data and compute. currently filtering more data - next run should have way fewer bugs/confabulations.
- Developed by: bleepybloops
- License: apache-2.0
- Finetuned from model : arcee-ai/Trinity-Nano-Preview
This afmoe model was trained 2x faster with Unsloth and Huggingface's TRL library.
- Downloads last month
- 34