Providers who deliver custom model fine-tuning.

Custom model fine-tuning is the practice of taking a pretrained base model and continuing training on a narrower dataset so its outputs shift toward a specific domain, style, or task distribution. In current practice this covers full-parameter fine-tuning, parameter-efficient adapters such as LoRA and QLoRA, instruction tuning on curated pairs, preference optimization (DPO/ORPO and successors), and distillation of a larger teacher into a smaller student. It is a class of systems and processes, data pipelines, training runs, evaluation harnesses, and serving infrastructure, not a single tool.

A buyer evaluating custom LLM fine-tuning should ask concrete questions. First, which base models are supported and how are licences handled, since a LoRA fine-tuning run on a restrictively licensed base can leave the resulting weights unusable in production. Second, how is the training set constructed, deduplicated, and held out, a real evaluation set that predates training is the single strongest signal. Third, what happens after the run: does the provider ship weights, an adapter, or only an API endpoint; how is drift monitored; and what does retraining cost when the base model version changes.

Common failure modes in model fine-tuning services are worth naming plainly. Vendors sometimes present retrieval-augmented prompting as fine-tuning, when no weights are updated at all. Small LoRA runs on a few hundred examples are often oversold as "custom models" while producing regressions on general capability. And distillation pitches routinely omit the teacher's licence terms, which govern whether the student can be used commercially.

SERPs for this category are hype-heavy, which makes independent verification hard. Trustgent grades providers on the L0-L5 evidence spectrum described at /how-we-verify, moving up only as reproducible training artefacts, held-out evaluations, and client-confirmed deployments accumulate.