This is a test model because the previous attempt failed. So it turns out I also trained this incorrectly due to the dataset using wrongly formatted ShareGPT so after I trained it again "correctly" for another full epoch I noticed that so.. I will have to train it yet again.. but at least now it should be fixed. But this model is using the wrongly formatted dataset. The next model should be better.
Trained with regular LoRA (not quantized/QLoRA) and LoRA rank was 128 and Alpha set to 32. Trained for 5000 steps (0.11 epoch).
This llama model was trained 2x faster with
Unsloth and Huggingface's TRL library.
