LoRA Training on RTX 5070: Exact Settings That Work

LoRA v15 training config for RTX 5070 (8GB):

Base: Qwen2.5-3B-Instruct-abliterated. Rank 32, alpha 64. LR 1e-4. Batch 1, grad accum 16. 2 epochs, fp16. 3,490 examples, 437 steps, 7h53min. Train loss 1.75, eval loss 1.04.

Key lesson: gradient accumulation 16 is critical. Without it, personality doesn't stabilize. With it, smooth convergence.

Conversion: convert_lora_to_gguf.py from llama.cpp. base_model_name_or_path must be HuggingFace name, not local path.

Drop your training configs.