nolabs-ai/deepfabric

What it solves

DeepFabric addresses the challenge of creating high-quality, diverse synthetic training data for language models, specifically for agentic behaviors like tool-calling and reasoning. It prevents model overfitting and hallucinations by ensuring datasets are domain-anchored and based on real tool execution results rather than simulated outputs.

How it works

The system uses unique topic graph generation algorithms to map out a domain and ensure comprehensive coverage of subtopics without redundancy. It then generates training samples using a ReAct (Reason-Act-Observe) loop. To ensure accuracy, it integrates with the Spin Framework to execute tools in isolated WebAssembly sandboxes, meaning the model's "observations" are based on actual state changes in a virtual filesystem. The generated data can be exported to Hugging Face and used with training frameworks like TRL, Unsloth, or Axolotl. Finally, it includes a built-in evaluation engine to test the fine-tuned model's tool selection and parameter accuracy on unseen tasks.

Who it’s for

It is designed for developers and AI researchers who are fine-tuning LLMs to act as agents, specifically those needing to teach models how to use tools, follow strict schemas, and perform complex planning in specific domains.

Highlights

• Topic Graph Generation: Uses graphs and trees to ensure 100% topic coverage and avoid redundant samples.
• Real Tool Execution: Integrates with WebAssembly (Spin) to provide authentic tool-calling traces instead of hallucinated simulations.
• MCP Support: Can import tool definitions from Model Context Protocol (MCP) server schemas.
• Integrated Evaluation: Provides a dedicated engine to measure tool selection accuracy, parameter accuracy, and execution success rates.
• Training Pipeline Integration: Direct compatibility with Hugging Face and popular SFT (Supervised Fine-Tuning) frameworks.