LoRA-JP — Japanese Domain LoRA Fine-tuning Pipeline

Challenge

Generic LLMs handle internal abbreviations, in-house terminology, and document formats poorly, and prompt engineering alone can't close the gap. Full fine-tuning, on the other hand, is too VRAM- and cost-heavy for small in-house experiments.

Solution

Unsloth-driven QLoRA (4-bit quantization + LoRA) makes training feasible on a single A100 40GB. TRL's SFTTrainer and DPOTrainer combine supervised and preference tuning, with WandB tracking loss and sample outputs. Resulting adapters are hot-swapped through vLLM's dynamic LoRA loading.

Results

• 1.9× training throughput on identical hardware (Unsloth vs. vanilla HF)
• +14% task accuracy on the internal eval set (internal benchmark)
• Three adapters served concurrently via vLLM hot-swap with zero restarts
• Training-job configs unified into reviewable YAML in PRs