Supervised fine-tuning that teaches the model how your work is done.
SFT is the foundation of every custom model: a base model learns your formats, vocabulary and procedures from carefully curated examples. Done well, it closes most of the gap to production on its own.
Supervised fine-tuning trains a pre-trained language model on input–output pairs that demonstrate exactly the behavior you want. It is the most reliable way to teach a model a consistent output schema, a house style, a domain’s terminology, or a multi-step procedure such as reading a claim, extracting the relevant fields and flagging inconsistencies.
The method is well understood. The outcome is not, because it depends almost entirely on the data. A few thousand clean, diverse, expert-verified demonstrations beat a hundred thousand scraped ones. Most of our SFT work is therefore data work: choosing which examples to label, writing guidelines that make experts agree with each other, removing near-duplicates and contradictions, and balancing the long tail of rare but important cases.
Signs you need supervised fine-tuning.
The right technique, chosen by the eval.
- Full-parameter SFT. Maximum quality for a dedicated model; best when you will serve one model per task.
- LoRA / QLoRA adapters. Train 0.1–2% of parameters; serve dozens of task adapters on one base model with multi-LoRA serving.
- Continued pre-training. Adapt a base model to a large private corpus before instruction tuning — for deep domain vocabulary.
- Curriculum & packing. Ordering, mixing and sequence packing to stabilize training and cut GPU hours.
- Curated, versioned SFT dataset with data card
- Trained weights or adapters, plus training configs
- Held-out eval report vs. base model and frontier baseline
- Quantized serving build and inference benchmarks
All assets transfer to you at handover. No licence fees.
Questions buyers ask us
How many examples do we need for supervised fine-tuning?
For format and style, 500–2,000 high-quality examples often suffice. For a demanding domain task, 3,000–20,000 is typical. Quality and coverage matter more than raw count; we usually find that the first 2,000 carefully chosen examples deliver most of the gain.
LoRA or full fine-tuning?
LoRA is cheaper to train and lets you serve many adapters on one base model; full fine-tuning tends to win by a point or two on hard tasks and on larger distribution shifts. We run both in the bake-off when the difference could matter and let the eval decide.
Will fine-tuning make the model forget general abilities?
Some forgetting is normal. We control it by mixing general instruction data into training, using conservative learning rates and LoRA where appropriate, and tracking a general-capability regression suite on every checkpoint.
Bring us the task. We'll bring the model.
A 45-minute technical briefing with a senior research engineer. You leave with a feasibility read, a data plan and a fixed quote — whether or not you hire us.
Book a technical briefing →