Complete model card: training details, eval, usage, limitations

V2 LoRA trained (loss 0.78, 3 epochs on augmented data)

initial commit