85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is fine-tuning?

Fine-tuning is further training of a pretrained language model on a smaller, task-specific dataset, updating its weights so it adopts a format, style or narrow skill that prompting alone does not hold reliably.

Prompting and retrieval change what a model sees on each request. Fine-tuning changes the model. A training job runs over example inputs and target outputs, or over preferred and rejected answers, and the resulting weights carry that behaviour into every later call without the examples in the prompt.

The decision has shifted in 2026. OpenAI's deprecations page records that from 7 May 2026 its self-serve fine-tuning stopped accepting organisations that had not fine-tuned before, and that from 6 January 2027 active existing customers can no longer create new fine-tuning jobs; models already tuned stay available until their base models are deprecated. For many teams, fine-tuning now means open-weight models on their own or a cloud provider's infrastructure.

Common methods

  • Supervised fine-tuning (SFT)

    Train on examples of correct responses. The standard method for formats, styles and narrow tasks.

  • Preference tuning (DPO)

    Direct preference optimisation trains on pairs of better and worse responses to the same prompt.

  • Reinforcement fine-tuning (RFT)

    A grader scores the model's responses and training reinforces the higher-scoring reasoning. OpenAI's guide lists it for reasoning models only.

  • LoRA

    Low-rank adaptation trains small added matrices instead of all weights. The LoRA paper (Hu and colleagues, 2021) reported 10,000 times fewer trainable parameters and about a third of the GPU memory compared with full fine-tuning of GPT-3 175B.

When fine-tuning earns its place

  • A fixed format at volume

    Extraction or classification run thousands of times a day, where shorter prompts and a smaller model cut cost.

  • A smaller model held to a baseline

    OpenAI's guide to agents advises setting an eval baseline on a capable model, then trying smaller ones. Tuning is one way to bring a small model up to it.

  • Style that examples do not hold

    A tone or domain convention that few-shot prompts reproduce inconsistently.

Where it goes wrong

  • Used for knowledge

    Facts trained into weights go stale, cannot be cited and cannot be restricted per user. Retrieval handles changing knowledge.

  • Locked to a base model

    A tuned model retires with its base model, and every upgrade means another training run and eval.

  • No eval to prove it

    Without a baseline on the untuned model, there is no way to show tuning helped.

How Motiontic's engineers apply it

Motiontic's engineers try context, then retrieval, then tuning, in that order, and keep whichever closes the gap on the client's eval set at the lower running cost. When tuning wins, the training data, base model and eval results are versioned together so the job can be repeated on the next model.

Fine-tuning methods compared

Fine-tuning methods compared
MethodTrains onFits
Supervised fine-tuningExample inputs with correct outputsFormats, styles, narrow tasks
Direct preference optimisationBetter and worse responses to the same promptTone and preference alignment
Reinforcement fine-tuningGrader scores on the model's own responsesReasoning tasks with checkable answers
LoRASame data as SFT, with small adapter weightsOpen-weight models on limited hardware

Common questions

What is fine-tuning an LLM?

Training an already-trained language model further on a smaller set of examples so it learns a specific format, tone or task, which changes the model's weights.

Is fine-tuning better than RAG?

They solve different problems. Retrieval supplies knowledge that changes or must be cited; fine-tuning shapes behaviour that stays fixed. Most teams should prove retrieval and prompting first.

Can OpenAI models still be fine-tuned?

Only by organisations already using the service. OpenAI closed self-serve fine-tuning to new organisations on 7 May 2026 and stops new jobs for existing customers on 6 January 2027, while keeping tuned models available until their base models are deprecated.

What is LoRA fine-tuning?

A parameter-efficient method that trains small added weight matrices instead of the full model, cutting memory and storage needs. It is widely used to tune open-weight models.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.