LoRA, QLoRA, and SFT on your proprietary corpus. The base model's general intelligence stays intact - your domain knowledge gets layered on top. Faster, cheaper, production-ready.
| Method | Data needed | Compute cost | Best for | We use it when |
|---|---|---|---|---|
| SFT | 1k - 100k examples | Medium | Instruction following | You have labeled examples |
| LoRA | 500-50k examples | Low | Style & domain shift | Consumer GPU budget |
| QLoRA | 500-50k examples | Very low | Large models on small GPU | Llama/Mistral on A100 |
| DPO | Preference pairs | Medium | Preference alignment | You have human feedback |
Fine-tuning is worth doing when prompting has plateaued: the model understands the task but consistently misses your domain’s vocabulary, format or judgement calls. Modulus handles dataset curation, the training run (SFT, DPO, LoRA or QLoRA), benchmark evaluation against the base model, and deployment. Where retrieval would solve the problem more cheaply, we say so before taking the work — most requests for fine-tuning are better served by a well-built RAG pipeline. You own the resulting adapter weights outright. Modulus works with teams in the United States, United Kingdom, Singapore, Hong Kong, Australia, Indonesia, Germany and France.
Fine-tune when the model needs to learn a style, format or judgement pattern. Use retrieval when it needs access to facts that change. Most requests framed as fine-tuning problems are actually retrieval problems, and retrieval is cheaper to build and to maintain.
Supervised fine-tuning (SFT) for format and task adherence, DPO for preference alignment, and LoRA or QLoRA where parameter-efficient adaptation keeps training and inference costs down. The method is chosen against the task, not by default.
Fewer examples than most teams expect — often several hundred to a few thousand high-quality pairs. Data quality and consistency matter far more than volume; a small clean set outperforms a large noisy one reliably.
You do. Modulus delivers the adapter weights, the training data pipeline and the evaluation harness. There is no dependency on a Modulus-hosted endpoint to keep using what was built.
Every engagement includes a benchmark suite built before training starts, scoring the base model and the tuned model on the same held-out set. If the tuned model does not beat the baseline on your task, that result is reported rather than hidden.
Free discovery call. Adapter weights yours. 30-day post-launch support.
Tell us your model, dataset size, and target task.