Playbook: Choosing the Right Fine-Tuning Method for Your LLM
Last month I burned two weeks trying to decide between full fine-tuning, LoRA, and QLoRA for a customer-support chatbot. The research was crystal clear on how each works. It was useless on when to use them. That frustration is exactly why I'm sharing this playbook I've been using in production.
The Real Decision Tree (Not What Papers Tell You)
Most guides dump a technique list and call it a day. Here's what actually matters:
Transfer Learning — When you have zero labeled data and need something working yesterday. Fast, cheap, often good enough. Skip if you have domain-specific jargon or tone requirements.
Full Fine-Tuning — Every parameter updated. Maximum performance ceiling, maximum everything else: VRAM, time, cost. Only worth it if you have 50k+ quality examples and dedicated GPU budget.
LoRA — Low-rank adaptation. Freezes the base model, trains tiny adapter matrices. I use this as my default starting point. 90% of the gain for 10% of the resources.
QLoRA — LoRA + 4-bit quantization. Run 7B models on a 24GB card. Sounds great until you hit the accuracy drop on specialized domains. Test carefully.
Continued Pretraining — Don't confuse this with fine-tuning. You're not adapting to a task, you're adapting the base knowledge. Critical for heavy industry jargon or non-English corpora.
Knowledge Distillation — Teacher-student setup. Massive engineering overhead. Only worth it when you need to shrink a model for edge deployment and latency is a hard constraint.
What I Actually Check Before Picking
1. Dataset size < 1k? → Start with transfer learning
2. Need sub-50ms inference? → Consider distillation
3. VRAM < 2x model size? → QLoRA or LoRA only
4. Domain accuracy critical? → Full fine-tuning (if data allows)
5. Budget constrained? → LoRA first, alwaysThe GitHub Playbook That Saved My Sanity
There's a repo that structures this properly — not just technique descriptions, but trade-off matrices and decision flows. It walks through each method with real numbers (VRAM usage, training time, accuracy deltas) rather than theoretical comparisons.
I've been using it as my go-to reference when onboarding new team members. The chapters on Foundation Models and Continued Pretraining are particularly solid — exactly the context most tutorials skip.
More chapters are incoming on Agentic AI and Production Systems, which is where the rubber really meets the road.
Built by the WildOctopus team. If you're tired of research-code gap, this is the kind of resource that turns papers into actual engineering decisions.