It seems fine but I should probably add some instruction prompts to the dataset or train it with a instruct dataset first and then train it with the RP stuff to make it better.
Another thing to note is this was trained with regular LoRA (not quantized/QLoRA) so it should improve the quality a bit. This model's context length is only 4096 so it's trained on that too but I think you can use RoPE with it.
LoRA rank was 128 and Alpha set to the same. Trained for 1 epoch.
This llama model was trained 2x faster with
Unsloth and Huggingface's TRL library.
