DataDreamer: Generate Synthetic Data and Train LLMs

Overview

DataDreamer brings LLM prompting, synthetic-data generation, and model training into a Python workflow. It addresses the work of connecting these stages and making experiments easier to resume, reproduce, and share.

It is aimed at researchers and developers who need to build multi-step LLM workflows or create training data, then fine-tune, instruction-tune, distill, or align models. It supports open-source and API-based LLMs, so it can fit both model experimentation and data-preparation pipelines.

Key Features

• Build multi-step prompting workflows with open-source or API-based LLMs.
• Generate synthetic datasets for new tasks or augment existing datasets.
• Fine-tune, instruction-tune, distill, and align models using existing or synthetic data.
• Cache work and resume workflows to support more efficient iteration.
• Use quantization and parameter-efficient training techniques such as LoRA.
• Share workflows and publish datasets and models with generated data cards, model cards, and citation information.

Use Cases

• Researchers can run reproducible experiments that connect LLM prompting, data generation, and model training.
• ML engineers can create synthetic examples to bootstrap a dataset or augment existing training data.
• Teams can build instruction-tuning or distillation pipelines when they have data and a suitable model to train.
• Developers can prototype workflows that combine API-based and open-source LLMs.

Project Facts

• Language: Python
• License: MIT
• Stars: 1.1k
• Forks: 59
• Topics: alignment, deep-learning, fine-tuning, gpt, instruction-tuning, llm, llmops, llms, machine-learning, natural-language-processing, nlp, nlp-library, openai, python, pytorch, synthetic-data, synthetic-dataset-generation, transformers
• Archived: no

Getting Started

Install the package:

pip3 install datadreamer.dev

See the README and documentation for setup details and examples.

Alternatives

• LLMBox: LLMBox centers on unified LLM training and evaluation, while DataDreamer focuses on reproducible workflows for data generation, training, and alignment.
• LlamaFactory: LlamaFactory focuses on fine-tuning through CLI and web interfaces, while DataDreamer also provides workflow tools for synthetic data generation and research.
• torchtune: torchtune provides editable PyTorch recipes for model post-training, while DataDreamer offers broader resumable workflows for data generation and model training.
• RL4LMs: RL4LMs specializes in reinforcement learning from custom rewards, while DataDreamer supports broader LLM workflows, including synthetic data and multiple training approaches.

Considerations

• Training and alignment workflows require appropriate model and compute resources; the supplied project information does not specify hardware requirements.
• The library supports several workflow stages, so users should consult the documentation for compatible models, integrations, and configuration details.
• The repository is not archived. Its listed latest push was on 2025-02-02, which is a snapshot and does not establish current maintenance activity.