DataDreamer

DataDreamer is an open source Python library for prompting, synthetic data generation and model training workflows. Its authors describe it as simple, efficient and research grade. A single Python script can load a large language model, generate a synthetic dataset from an instruction, transform that dataset with further prompts, and then fine tune a smaller model on the result, with every step's output cached and saved to a working folder. The library is aimed at machine learning researchers and engineers who need reproducible pipelines rather than one off notebooks. The documentation walks through generating training data with attributed prompts, distilling a stronger model's capabilities into a weaker one, augmenting and cleaning existing datasets, bootstrapping few shot examples, instruction tuning a model, aligning a model with human preferences, and training a self improving model through self rewarding. Advanced guides cover caching and saved outputs, writing custom steps, LLM wrappers and trainers, running steps in parallel, running and training models across multiple GPUs, quantization, and parameter efficient training. The API reference spans datasets, embedders, LLMs, retrievers, steps, task models and trainers. DataDreamer is free to use. It installs from PyPI with a single pip command, and the source is published on GitHub. There is no paid plan or hosted service; the only costs are the compute and any commercial model APIs a pipeline calls. The documentation is versioned, and the release shown at the time of writing is v0.46.0, with a citation page provided for academic use.