Overview
Understand what Adaptive Data produces and how data moves from import through adaptation and evaluation.
Adaptive Data analyzes the structure of your source data, then adapts and optimizes it into a model-ready dataset for AI training.
What you can build
Section titled “What you can build”- Instruction datasets — prompts paired with expected completions for supervised fine-tuning
- Preference pairs — two completions compared to indicate which is better, for preference-based training
How it works
Section titled “How it works”Adaptive Data runs through six stages:
- Import — bring in data from a local file, Hugging Face, or Kaggle.
- Map columns — identify prompt, completion, context, image, or chat data.
- Apply recipes — optionally deduplicate or rephrase prompts and generate reasoning traces.
- Expand — optionally augment, translate, or localize rows into a new dataset.
- Set brand controls — optionally control length, safety, grounding, tone, and other Blueprint instructions.
- Adapt and evaluate — start adaptation, then compare quality signals for the source and adapted data.
After completing multiple runs, you can combine their datasets into one new training dataset.
Supported inputs
Section titled “Supported inputs”- File formats: CSV, JSON, JSONL, Parquet, PDF, DOCX, PPTX, XLSX, HTML, ZIP, and TXT
- Hugging Face: point at a dataset URL
- Kaggle: point at a dataset page URL
Start with the Adaptive Data quickstart, or browse the complete API reference.