Adapt a pretrained LLM to your domain with LoRA, QLoRA, or full SFT.

Fine-Tuning — Specializing Foundation Models

Fine-tuning continues training a base model on curated examples to teach style, format, domain knowledge, or new skills. Modern parameter-efficient methods (LoRA/QLoRA) update <1% of weights, training in hours on a single GPU at a fraction of full SFT cost.

Interactive Pipeline Flow

IngestProcessServe

📚 Curate — instruction pairs

Collect instruction-response pairs from production traces, SME-authored examples, or synthetic generation. Aim for 1k-10k diverse, high-quality samples.

Architecture Diagrams

End-to-end FT pipeline from data prep to deployed adapter.

Loading diagram...

Key Design Decisions

📊Data Quality > Volume

1,000 hand-crafted examples beat 100,000 noisy ones. LIMA paper showed 1k examples can match RLHF on style.

🪶LoRA is Default

Full SFT rarely needed. LoRA r=16 captures most domain adaptation; QLoRA fits 70B on a single 24GB GPU.

⚠️Catastrophic Forgetting

Aggressive FT degrades base model's general abilities. Always benchmark MMLU/HellaSwag pre/post.

🎭Mask System + User

Compute loss only on assistant tokens. Training to predict the user's question wastes capacity and hurts quality.

🔀Adapter Hot-Swap

Multi-tenant serving: keep base in GPU, swap LoRAs per request (vLLM, S-LoRA). Avoids per-tenant model duplication.

💰FT > RAG When

Use FT for tone, format, terminology. Use RAG for facts, freshness, citations. Most apps need both.

Discussion