Open source · Synthetic data

Build smaller, smarter models that punch above their weight

High-quality synthetic datasets with real tool execution, diverse topic graphs, and built-in evaluation.

Get startedView on GitHub
topic_graph.generate()
01 DOMAIN02 TOPICS03 SUBTOPICS04 SAMPLESKubernetesdomainNetworkingtopicStoragetopicSecuritytopicIngressServicesDNSVolumesClaimsCSIRBACSecretsPolicies

  1. 01 DOMAIN

    Name the field of expertise you want a model to master.

  2. 02 TOPICS

    The domain is split into its major areas of knowledge.

  3. 03 SUBTOPICS

    Each area expands into specific, non-overlapping subtopics.

  4. 04 SAMPLES

    Every subtopic yields training samples, with tools actually executed.

Architecture

How the DeepFabric pipeline works

  1. 01

    Domain input

    Start from a topic of expertise.

  2. 02

    Topic graph

    Coverage without redundancy.

  3. 03

    Training samples

    Generated with real tool execution.

  4. 04

    Evaluation

    Prove the capabilities you trained.

Why DeepFabric

Domain-anchored diversity without redundancy

Diverse topic graphs

Unique topic graph generation covers your domain thoroughly while avoiding the repetitive samples that cause overfit.

Real tool execution

Tools actually run, so samples are grounded in real outputs instead of simulated ones.

Built-in evaluation

Evaluate on unseen tasks derived from your training splits.

Export anywhere

Send datasets straight to TRL, Unsloth or Axolotl and start fine-tuning.

Standard approach

  • Repetitive samples
  • Simulated tool outputs
  • No validation

Models overfit and hallucinate.

DeepFabric approach

  • Diverse topic graphs
  • Real tool execution
  • Strict validation

Models are grounded and efficient.

Tool execution

Real tool calls, mocked safely with WebAssembly

Simulated tool outputs teach models to hallucinate. DeepFabric runs your tools in a Spin sandbox instead: each one is a WebAssembly component backed by mock data, so the agent gets genuine, deterministic results without touching a live API.

LLM agentplans + calls toolsWEBASSEMBLY SANDBOX · SPINread_filewasm componentwrite_filewasm componentyour_toolfrom your MCP schemamock data + seeded filestool callresultloggedTraining samplereal tool results

  1. 01 IMPORT

    Pull tool schemas from your MCP server into the sandbox.

  2. 02 MOCK

    Load mock responses and seed files for the tools to work with.

  3. 03 EXECUTE

    Agent calls run as WebAssembly components, so they really execute.

  4. 04 RECORD

    Real results are logged into each training sample.

Training

Train

Turn your dataset into a fine-tuned model with the frameworks you already use.

Dataset.jsonl samplestrl_sftformatterSFTTrainerTRL · Unsloth · AxolotlTuned modelknows tool format

The built-in trl_sft formatter turns generated samples into tool-calling conversations ready for TRL's SFTTrainer, or export to Unsloth and Axolotl. Then check the result with built-in evaluation on unseen tasks.

Open source and community driven

Built in the open. Contribute, share datasets, and help make small models more capable.

Star on GitHubRead the docs