This was made using the Llama 3 Instruct prompt formatting so that it should be easier to be merged with other models using that format.

LoRA trained in 4-bit with 8k context using meta-llama/Meta-Llama-3-8B as the base model for 1 epoch.

Dataset used is a modified version of grimulkan/LimaRP-augmented.

There might be a slight issue with the prompt formatting since Unsloth decided to leave "gpt" and "user" to the prompts which seem to not be handled correctly all the time.

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.