Resume training from step 200 checkpoint

Fix: add formatting_func for Unsloth SFTTrainer"

Unsloth + bf16 LoRA for Qwen3.5-9B, unbuffered logs, proven working approach

Rewrite: QLoRA 4-bit frozen weights + bf16 LoRA, 5k examples, 1 epoch, fast saves, unbuffered logs

Fix: add health check server so Space doesn't timeout after 30min

Fix: HF_TOKEN handling, torch_dtype->dtype, trackio run_name->name

Upgrade: LoRA r=64, batch_size=2 for L40S 48GB VRAM

Fix: remove 4-bit quantization, use full bf16 + 8-bit optimizer for quality

Add DM LoRA training script