Use cases
16 tasks worth fine-tuning for.
What is fine-tuning actually good for?
Tasks where the decision boundary is yours rather than general — your taxonomy, your schema, your conventions, your definition of a good answer. Each page below gives the model sizes to start from, the dataset shape, the metric that decides success, and the mistake that most often ruins it.
What these have in common
Every case below shares one property: the thing that makes the task hard is specific to the organisation doing it. Which queue a ticket belongs in, what your schema calls a field, which table was deprecated last year, what your brand means by confident but not arrogant. A frontier model has to be told all of that on every request and still guesses at the edges, because it has never seen your edge cases.
That is the shape of problem fine-tuning solves well. The shape it solves badly is the opposite one: facts that change, tasks with no consistent right answer, and anything where the requirement is breadth rather than depth.
Each page gives the model sizes to start from in order, the dataset shape, the metric that decides success, and the mistake that most often ruins the project. The mistakes are worth reading even if the use case is not yours — most of them generalise.
Customer support triage
SFTTrain a small model on your own resolved tickets, labelled with the queue they ended up in.
Structured data extraction
SFTFine-tune on your own document-to-schema pairs.
Replacing a frontier API on one task
SFTCapture real production traffic, fine-tune a small model on it, then gate the candidate against the model it would replace on your own cases.
Domain question answering
SFTUse retrieval for the facts and fine-tuning for the behaviour.
Code review assistant
SFTTrain on your merged pull requests paired with the review comments that were acted on.
SQL generation over your own schema
SFTTrain on question-and-query pairs from your actual query logs.
Tool-calling agents
SFTTrain on traces of successful tool sequences from your own system.
Content moderation
SFTTrain on your own moderator decisions, including the appeals that were overturned.
Summarisation in a house style
SFTTrain on a few hundred summaries your team actually approved.
PII redaction
SFTTrain on your own documents with your own redaction decisions.
Reasoning with reinforcement learning
GRPOUse reinforcement learning when you can score an answer automatically but cannot write the ideal reasoning that produced it.
Translation and localisation
SFTTrain on your own translation memory.
Classification at volume
SFTFine-tune the smallest model that clears your accuracy bar and serve it on an entry GPU class.
On-brand copy generation
DPOTrain on copy your brand team approved, then use preference tuning on pairs they ranked.
Semantic routing between models
SFTFine-tune a very small model to predict whether the large one is needed.
Document parsing pipelines
SFTTrain on your own extracted text paired with the corrected output.
Objectives at a glance
- SFT — Supervised fine-tuning, used by 14 of the cases above.
- DPO — Preference tuning (DPO), used by 1 of the cases above.
- GRPO — Reinforcement learning (GRPO), used by 1 of the cases above.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.