SFT Studio Pro is a workspace for building supervised fine-tuning (SFT) datasets and moving them through a connected workflow: complete project onboarding, create or convert examples, review batches, assemble the approved corpus, train a model, and try it with a free model chat sample.
This manual follows the order in which the work is completed in the application. It includes SFT background, dataset planning, project onboarding and definition, automated and manual batch generation, the Document Converter, the optional FTK TRiAD alignment corpus, QA, the Dataset Registry, Automated Training, and model access.
Understanding SFT
Supervised fine-tuning is teaching by example. A model is shown examples of inputs and the responses desired for those inputs. Across the training examples, the model learns patterns in format, voice, content, and response style. An SFT example may be an instruction-response item or a multi-turn conversation, depending on the project's selected example type.
The completed dataset is a collection of structured examples. SFT Studio Pro organizes these examples in batches so they can be generated or imported, reviewed, and approved before being added to the project corpus.
Planning the dataset
There is no single correct dataset size for every project. The amount of data depends on the model, the range of tasks, and the intended identity and capabilities of the finished model. A focused assistant with a narrow purpose can be well served by a smaller, carefully reviewed corpus, while a model expected to handle varied requests benefits from broader coverage.
Quality matters more than raw count. A modest set of accurate, consistent, well-structured examples usually outperforms a larger set that is uneven. Because the app builds the dataset in batches, you can start with a core set of topics and grow the corpus incrementally — reviewing each batch as it arrives rather than committing to a final size up front.
Quality and coverage
A dataset can contain many examples and still cover only a narrow range of situations. Coverage means the corpus spans the skills, subjects, and interaction types the model will face — including varied phrasings, different difficulty levels, short and extended exchanges, and ambiguous requests.
Quality is enforced at two points. First, the golden example approved during onboarding defines the structure and standard every generated example follows. Later, Quality Assurance analyzes each generated or converted batch, flags potentially problematic examples with written explanations, and leaves the final approval decision to you.
Training Corpus and Validator
The batches you select in the Dataset Registry compile into the project corpus. During Compile & Export, the app automatically partitions the selected examples into two files: a training set (the bulk of the examples) and a validation set (approximately 10%), used to measure the model during training.
The Automated Trainer takes both of these splits, verifies that they are non-empty and match the required ChatML format, and runs the training pipeline once you approve and submit the fine-tune job. The validator is what makes the training run measurable: it holds back examples the model did not train on, so its progress is tested against unseen data.