Fine-Tuning Experiments (LoRA/QLoRA)

Semi-hands-on configs and scripts for fine-tuning open-weight LLMs. Nội dung bán thực hành: config & script cho fine-tuning LLM open-weight.

⚠️ Work in progress. Small configs first, real runs later.

Goals

• Minimal LoRA/QLoRA config with PEFT.
• Reproducible train/eval commands.
• Keep GPU memory modest (8–24GB target).

Layout

• configs/ – YAML configs (model, LoRA ranks, train args).
• scripts/ – train/infer helpers.
• data/ – tiny demo JSONL (placeholder).
• logs/ – example training logs (truncated).

Quick Start (conceptual)

# install
pip install transformers peft accelerate datasets bitsandbytes

# dry-run (conceptual)
python scripts/train_lora.py --config configs/lora-llama3-8b.yaml