DeepFabric
A tool for generating synthetic training data for language models and agent evaluations, combining reasoning traces with tool-calling patterns through topic-graph-guided sample generation.
DeepFabric generates synthetic training data for language models and agent evaluations by combining reasoning traces with tool-calling patterns, aiming to teach models to plan, act, call tools correctly and follow strict schemas. A topic-graph generation algorithm spreads sample creation across subtopics to keep a dataset diverse and reduce the redundancy that can cause a model to overfit.
Constrained decoding, response validation, and real tool executions inside isolated WebAssembly sandboxes keep generated samples matched to a schema; tool definitions can be imported directly from MCP server schemas and mocked automatically. It can be used as a library, a CLI tool, or through YAML configuration, generated datasets can be uploaded to Hugging Face and imported into training frameworks such as TRL, Unsloth and Axolotl, and a built-in evaluation engine scores a trained model on held-out tasks.