Дообучение LLM через LoRA и PEFT

ai-tooling

Fine tuning expert is a Claude Code skill that guides you through fine-tuning large language models using LoRA and QLoRA adapters, Hugging Face PEFT methods, instruction tuning, RLHF, DPO, and post-training quantization. Reach for it when you need to adapt a foundation model to a specific domain — say, a support chatbot with a particular tone or a document classifier — when GPU memory is tight and you need 4-bit QLoRA to fit training into a single card, or when you want to run a fine-tuning job via the OpenAI fine-tuning API without managing your own infrastructure. The skill walks through a structured workflow: validating and formatting JSONL datasets, selecting a PEFT method based on available memory, configuring hyperparameters (learning rate, batch size, cosine scheduler, warmup ratio), monitoring train/validation loss, evaluating with perplexity and task-specific metrics like BLEU and ROUGE, then merging adapter weights into the base model for serving. Built on Python with transformers, peft, trl, and bitsandbytes; ready-to-run LoRA and QLoRA code examples are included. The focus is training and adaptation, not inference orchestration — for production serving pipelines, pair it with the devops-engineer skill.