Deepfabric

DeepFabric is an open source Python toolkit for generating synthetic datasets used to train and evaluate language models. A user supplies a root topic and a system prompt, and DeepFabric builds a topic graph that recursively expands into subtopics, then generates training samples for each node so the dataset covers the subject broadly without repetitive examples that encourage overfitting. It is aimed at machine learning engineers fine tuning smaller models for specific domains or for tool use. Samples can be generated as chain of thought conversations with reasoning traces, and tool calls run in sandboxed environments rather than being simulated, so the data reflects real tool behavior. Constrained decoding and strict validation keep output syntax and structure correct. Generation works with OpenAI, Anthropic and Google Gemini through their API keys, or entirely locally through Ollama with no key required. Beyond generation, optional extras cover training and evaluation, and the command line interface can upload finished datasets to Hugging Face or Kaggle, import tool definitions and report configuration details. DeepFabric is free and installs from PyPI with pip or uv, or from source on GitHub. No pricing is published; costs come only from the model provider used for generation. The source is maintained on GitHub, and the listed address is the project documentation site, which also covers the training and evaluation utilities in detail.