Model save

Fix: add assistant_only_loss=False to prevent all labels being masked to -100

Training in progress, step 500

Add adapter resume support: loads previous adapter from hub if available, 48h timeout ready

Training in progress, step 500

Optimal config: high-quality Hebrew data, constant LR, packing disabled, lora_dropout=0.1

Switch to packing=True, batch=4, grad_acc=4, step-based checkpoints for speed

Disable eval to fix OOM on A10G

initial commit