This model was trained with SFT using Unsloth on the ChatML format with 8k context. Carp models are trained with a combination of pretrain, instruct, and chat datasets.
This qwen2 model was trained 2x faster with
Unsloth and Huggingface's TRL library.
