Skip to content

Latest commit

 

History

History
49 lines (40 loc) · 4.25 KB

File metadata and controls

49 lines (40 loc) · 4.25 KB

Fine-Tuning Infrastructure

Concept

  • Fine-tuning adapts a pre-trained model (especially an LLM or foundation model) to a specific task, domain, or style by further training it on a smaller, targeted dataset — leveraging the base model's general knowledge instead of training from scratch.
  • The spectrum of adaptation, from cheapest to most involved:
    • Prompting / in-context learning — no training; just instructions/examples in the prompt (topic 9). Cheapest, most flexible.
    • RAG — inject retrieved knowledge at inference (topic 7); good for knowledge, not behavior.
    • PEFT (Parameter-Efficient Fine-Tuning) — e.g., LoRA/QLoRA: train small adapter weights while freezing the base model. Drastically cheaper than full fine-tuning (tiny fraction of parameters), and you can swap adapters per task.
    • Full fine-tuning — update all model weights. Most powerful but most expensive and storage-heavy.
    • RLHF / preference tuning (DPO) — align the model to human preferences.
  • Fine-tuning teaches behavior, format, and style (and narrow domain skills); it's generally not the right tool for injecting frequently-changing facts (use RAG for that).
flowchart LR
    BASE[Pre-trained base model] --> METHOD{Adaptation method}
    METHOD -->|prompt/RAG| NOTRAIN[No weight changes]
    METHOD -->|LoRA/QLoRA - PEFT| ADAPT[Small adapter weights]
    METHOD -->|full FT| ALL[All weights updated]
    ADAPT --> SERVE[Serve base + adapter]
Loading

Problem It Solves

  • Gets a model to reliably do a specific task/format/tone that prompting alone can't achieve consistently — e.g., a consistent structured output, a specialized domain style, or a narrow classification — at lower inference cost than a huge prompt.
  • PEFT/LoRA makes adaptation affordable: fine-tune a large model on modest hardware by training a tiny set of adapter parameters, and serve many task-specific adapters over one shared base.
  • Reduces reliance on long, expensive prompts by baking behavior into weights.

Trade-offs

  • Fine-tuning vs. RAG vs. prompting (the key decision) — prompting/RAG are cheaper, faster to iterate, and keep knowledge current (no retraining to update facts); fine-tuning changes behavior/style but bakes in knowledge that goes stale and requires retraining to update. Use RAG for knowledge, fine-tuning for behavior — confusing the two is a common, costly mistake.
  • PEFT vs. full fine-tuning — LoRA/QLoRA is far cheaper (compute, storage, swappable adapters) and usually sufficient; full fine-tuning is more powerful but costly and produces a full model copy per task.
  • Cost & data needs — fine-tuning needs a quality curated dataset and compute; poor data makes it worse, not better. Quality > quantity.
  • Maintenance — a fine-tuned model must be re-tuned as the base model upgrades or requirements change; it's an ongoing commitment, not one-and-done.
  • Catastrophic forgetting / overfitting — over-fine-tuning can degrade the base model's general abilities or overfit the small dataset.
  • Evaluation is essential — you need evals (topic 26) to know fine-tuning actually helped and didn't regress other behaviors.

Examples

  • LoRA adapters per task
    • One base LLM serves multiple products via different LoRA adapters (support tone, code style, summarization format) — swapped at inference, far cheaper than separate full models.
  • RAG vs. fine-tune decision
    • "The model should answer from our latest docs" → RAG (docs change). "The model should always reply in our brand voice and a fixed JSON schema" → fine-tune (behavior).
  • QLoRA on modest hardware
    • Quantized base + LoRA adapters let a large model be fine-tuned on a single GPU, democratizing adaptation.
  • Preference tuning
    • DPO/RLHF aligns outputs to human-rated preferences for helpfulness/safety, evaluated against a held-out preference set.
  • Interview framing
    • When adaptation comes up, lead with the decision tree: prompt → RAG → PEFT (LoRA) → full fine-tune → preference tuning, choosing the cheapest that works. The crisp rule "RAG for knowledge, fine-tuning for behavior/style," plus PEFT/LoRA for cost-efficiency and the need for evals, is exactly the modern, cost-aware LLM-systems signal.