Training Method: 16-bit LoRA with rank 64 / alpha 64. No rslora or loftq.
Validation Loss: ~0.18, down from 0.22 at 300 steps. I split 5% (761 samples) as validation, other 95% (14,448 samples) as training. Val loss was plateauing so I stopped at 2,104 steps.
Training Steps: ~2,100 (about 2.32 epochs with 16 batch size)
Training Ctx Length: 8,192 (no samples were longer than this)
