Skip to content
Multigrid

Fine-tuning & post-training

When changing a model's weights is the right answer, what each method actually costs, and the failure modes that make most fine-tunes a waste of a fortnight.

Fine-tuning is the most over-recommended intervention in applied machine learning. It is also, for a narrow set of problems, the only thing that works. The distance between those two sentences is where this cluster lives.

The pages start with the decision — whether to train at all — because that is where most of the money is lost, and only then get into method, data, evaluation and serving. Where a claim comes from a paper it is named. Where a number is derived, the derivation is shown so you can substitute your own model and redo it. Nothing here reports an experiment we ran, because we did not run one.

Should You Fine-Tune? A Decision Tree

A four-rung ladder — prompt, few-shot, retrieval, stronger model — and the test that tells you whether fine-tuning is the rung you are actually on.

5 min read

Fine-Tuning vs RAG: They Solve Different Problems

One changes what the model knows at inference time, the other changes how it behaves — and each mistake has a distinctive smell.

5 min read

LoRA Explained: Fine-Tuning Without the GPU Bill

Low-rank adaptation worked out on actual matrices: what the parameter counts are, why the memory saving is bigger than the parameter saving, and where it breaks.

6 min read

QLoRA and Fine-Tuning on One Consumer GPU

The 24 GB memory budget worked out term by term, what NF4 and double quantisation actually buy, and the tensor that causes most out-of-memory errors.

6 min read

How Much Data Do You Need to Fine-Tune?

Three published datasets that bracket the range, why the curve flattens so early for behaviour tasks, and a procedure that answers the question for your task instead of in general.

5 min read

Building a Fine-Tuning Dataset From Production Logs

Which requests to select, how to clean them without flattening the distribution, and the two leakage traps that make a fine-tune look better than it is.

5 min read

Supervised Fine-Tuning, Step by Step

What SFT actually optimises, an annotated configuration with a reason for every value, and how to read the loss curve you get back.

6 min read

RLHF Explained Without the Jargon

Three stages, one reward model, and a KL penalty — what each piece is for, written out with the actual loss and the failure it exists to prevent.

5 min read

DPO vs PPO vs GRPO: Preference Tuning Compared

Three ways to turn preferences into weights, compared on the thing that decides between them: how many models must be resident and whether training needs online sampling.

5 min read

Catastrophic Forgetting: What Fine-Tuning Breaks

Why training on a narrow task degrades everything else, what the literature reports about it, and the retention suite that catches it before your users do.

5 min read

Evaluating a Fine-Tuned Model Against the Base

Four evaluation suites rather than one, how many examples you actually need to detect the difference you care about, and why most fine-tuning evaluations cannot detect anything.

5 min read

Serving a Fine-Tuned Model: The Part Nobody Budgets For

Why a fine-tune moves you from per-token billing to per-hour billing, the break-even calculation between them, and the costs that are not compute.

5 min read

Multi-LoRA Serving: Many Adapters, One Base Model

How thousands of per-customer adapters share one base model in memory, the arithmetic that makes it work, and where the arrangement stops working.

5 min read

Distillation: Teaching a Small Model From a Big One

Two different techniques share the name — logit matching and training on generated text — and only one of them is available through an API.

5 min read

Continued Pre-Training for Domain Adaptation

When SFT is the wrong instrument because the corpus is enormous and unlabelled — with the compute budget derived so you can see the gap before committing to it.

5 min read

Synthetic Data for Fine-Tuning: Recipes and Risks

Four published generation recipes, what the model-collapse research actually showed, and the distinction between replacing real data and accumulating alongside it.

5 min read

Fine-Tuning for Format vs Fine-Tuning for Knowledge

Why the same technique works startlingly well for one of these and badly for the other, what the published evidence says, and a test that tells you which you have.

5 min read

The Real Cost of a Fine-Tune, End to End

A four-term cost model you fill in with today's numbers — data, compute, evaluation, serving — and the reason the compute term is almost never the one that matters.

6 min read

Licensing: What You May Legally Fine-Tune and Deploy

The four separate licences that apply to any fine-tune, what the major open-weight licence families actually require of derivatives, and the terms that most often surprise people.

6 min read

Fine-Tuning Failures: The Modes and How to Tell Them Apart

A taxonomy of documented fine-tuning failures — the crashes, the silent ones and the ones that look like quality problems — each with the diagnostic that identifies it.

6 min read

Other topics