How to choose an LLM fine-tuning platform or service

01

Decide whether fine-tuning is the right tool

Fine-tuning helps with consistent format, tone, tool use, narrow task accuracy, and moving a task to a smaller, cheaper model. It does not keep a model current with new facts; retrieval does that. Write down the task, the current baseline score on a held-out test set, and the target. If prompting and retrieval already meet the target, you may not need to train.

02

Pick the training method

• Supervised fine-tuning (SFT): prompt and ideal-response pairs. The usual starting point.
• LoRA or QLoRA: trains a small adapter instead of all weights. Lower cost, fast to swap, and often enough.
• Full fine-tuning: updates every weight. More control, more compute, and a full-size model to host.
• Preference tuning (DPO, ORPO): pairs of preferred and rejected responses. Good for tone and quality judgments.
• Reinforcement fine-tuning (RFT, RL): a grader or reward function scores outputs. Fits tasks you can check but cannot write one ideal answer for.
• Distillation: a larger teacher model generates training data for a smaller student model.

Check method support per model. Several vendors support a method on some base models but not others.

03

Prepare the data

Most managed APIs take JSONL chat-format files with one example per line. Keep separate training, validation, and final test sets. Price is usually based on training tokens, which equal dataset tokens times epochs, so estimate both before you start. For preference tuning you need chosen and rejected responses; for RL you need prompts and a reliable grader that you have tested against wrong and adversarial answers.

04

Choose the base model and settle weights ownership

Closed models (Gemini on Google Cloud, OpenAI models on Microsoft Foundry, Amazon Nova on Bedrock) are tuned and served inside that cloud; the weights stay with the provider. Open-weight models can often be downloaded after training, as merged weights or a LoRA adapter. If you need to move the model to your own infrastructure or another vendor later, confirm the download path and the license terms in writing before training.

05

Choose a buying model

• Managed fine-tuning API: upload data, set hyperparameters, the vendor runs the job. Fastest path.
• Training API or bring-your-own-code platform: you write the training loop or container; the vendor runs GPUs. Most control, most engineering.
• Hyperscaler cloud platform: tuning inside AWS, Google Cloud, or Azure accounts you already govern.
• Managed training service: a firm prepares data, trains, evaluates, and hands over a model. Least in-house work; contract terms matter most.

06

Plan deployment after training

Ask where the tuned model runs and how it is billed: serverless per-token, dedicated endpoints billed per hour or minute even when idle, or provisioned throughput. Some vendors serve LoRA adapters on shared capacity; others require a dedicated deployment for any trained model. Check tuned-model inference prices, which can differ from the base model.

07

Security, compliance, and cost controls

Ask for current reports, not badges: SOC 2 Type II report, ISO/IEC 27001 certificate and scope, HIPAA BAA if you handle health data, GDPR data processing terms, data residency for training as well as serving, and data retention and deletion for uploaded datasets and checkpoints. Set a budget per experiment, including minimum charges, storage, evaluation runs, and hosting.