Writing

Fine-tuning is a last resort, not a first move

Teaching a model is a five-rung ladder. Most teams skip to the expensive rungs. You should choose the shallowest intervention that solves your problem under evaluation.

"We should just fine-tune it" is the sentence that ends productive architecture conversations. Sometimes it is correct. Often it is a costly solution to the wrong problem. Teaching a model is not one technique: it is a ladder of interventions with different costs, different guarantees, and different failure modes. You should choose the shallowest rung that solves your problem under evaluation.

The most common expensive mistake is fine-tuning to add facts that should have been retrieved. The second most common is prompt-hacking to compensate for a capability that the model simply does not have.

The ladder

Five rungs, from cheapest to deepest: prompts with zero training, few-shot examples in context, retrieval and tools at inference time, parameter-efficient adapters like LoRA, and full fine-tuning that rewrites the weights. Higher rungs cost more in data, compute, and operational burden. They also reach deeper: the behavior persists without extra context. And they are harder to roll back.

The key distinction that most teams miss is between how and what. Fine-tuning changes how a model answers. Retrieval changes what it answers with. If your failure is "it doesn't know our internal policy," that is almost always a retrieval problem. If your failure is "it can't reliably follow a five-step procedure even when given the right facts," that is a capability problem, and the ladder starts to apply.

Pick a real symptom, watch it route to a rung

The first question is always "what" vs "how": is this missing knowledge, or a missing capability?

A "WHAT" PROBLEM: missing knowledge
PromptsFew-shotRAG / toolsLoRA / adaptersFull fine-tune
This is a knowledge gap, not a capability gap: a "what" problem. The model reasons fine; it just needs the current policy in context. Fine-tuning this in creates an update treadmill for a fact that changes.
Notice that "how" problems don't all land on the same rung either: tone is shallow, broken procedure adherence is deeper, and a default behavior baked into every response is deeper still. The ladder isn't a single jump from prompting to fine-tuning; it's a sequence, and the right rung is the shallowest one that actually matches the symptom.

This is why the LoRA rung exists at all: Hu et al. (2021) showed it can cut the number of trainable parameters by up to 10,000x and GPU memory requirements by 3x relative to full fine-tuning of GPT-3 175B, while matching or beating full fine-tuning quality on the benchmarks they tested. That cost gap is the entire argument for trying adapters before a full fine-tune.

The fine-tuning ladderChoose the shallowest rung that solves the problem under evaluation. Measure before climbing.
RungBest forDo not use when
PromptsFormat, tone, constraints the model already supportsThe capability is not there; prompts cannot install it
Few-shotComplex schemas, output patterns, structural imitationExamples are noisy or cover more than one task shape
RAG / toolsFresh facts, org-specific truth, citationsSource quality is poor or retrieval is unmonitored
LoRA / adaptersPersistent style, consistent tool-use patterns, domain registerThe knowledge changes frequently; adapters are not a database
Full fine-tuneDeep domain shift, default behavior everywhereYou lack eval harness; you will ship regressions invisibly
Operational cost vs behavior persistenceHigher rungs buy persistence at the cost of update agility. Neither extreme is universally right.
Loading chart...

Anti-patterns worth naming explicitly

Fine-tuning to memorize documentation is the most common waste of compute in applied AI. If the content changes weekly, baking it into weights creates an update treadmill, destroys auditability, and costs orders of magnitude more than a retrieval index update. The second anti-pattern is skipping rungs: jumping to fine-tuning before clarifying requirements, building an eval set, or checking whether the base model with better prompting already meets the bar. Training without evals is expensive guessing.

The rungs stack. The best production systems combine a fine-tuned model for style and tool-use patterns, retrieval for facts, and a short prompt for constraints. Climbing to the right rung means measuring at each step, not optimizing for the most impressive-sounding solution.

Which rung actually fitsThree questions that route you to a rung faster than a generic decision tree.

Changing facts should stay outside model weights; that's an update treadmill, not a fine-tune.

This is the same decision tree experienced practitioners reach for:

Akshay's prompting vs. RAG vs. fine-tuning decision framework.

References: LoRA: Hu et al. (2021), Hugging Face PEFT, Anthropic