DeepFabric

Getting started

1. 1

   DeepFabric also uses YAML configuration with 3 main sections and optional shared LLM defaults.

2. 2

   The YAML configuration example uses mocked tool execution and requires a Spin service, which the project provides in a docker image.

3. 3

   To use the YAML configuration, first save it as config.yaml, then run generation with that file as the input; this page does not include the YAML example, so follow the official documentation for its exact form.

4. 4

   Tool tracing can be enabled in the YAML config; the YAML code block on this page is only the tool tracing fragment, not a complete configuration, so it must be merged into a full YAML configuration with its three main sections before use, and the full configuration has to come from the official documentation rather than saving this fragment alone as config.yaml to run.

5. 5

   DeepFabric returns standard HuggingFace datasets.

6. 6

   Once a dataset is generated, it can be automatically uploaded to Hugging Face.

7. 7

   A generated dataset can be directly imported into training frameworks like TRL, Unsloth, and Axolotl.

pip install deepfabric

export OPENAI_API_KEY="your-api-key"

deepfabric generate \
  --topic-prompt "Python programming fundamentals" \
  --generation-system-prompt "You are a Python expert" \
  --mode graph \
  --depth 3 \
  --degree 3 \
  --num-samples 9 \
  --batch-size 3 \
  --provider openai \
  --model gpt-4o \
  --output-save-as dataset.jsonl

deepfabric generate config.yaml

generation:
  conversation:
    type: cot
    reasoning_style: agent

  tools:
    spin_endpoint: "http://localhost:3000"  # Spin service URL
    available:                              # Filter to specific tools
      - read_file
      - write_file
      - list_files
    max_agent_steps: 5                      # Max ReAct iterations

    # Optional: Seed initial state for scenarios
    scenario_seed:
      files:
        "config.json": '{"debug": true}'