Home
Synthetic Training Data for Agents
Quick Start¶
pip install deepfabric
export OPENAI_API_KEY="your-key"
deepfabric generate \
--topic-prompt "DevOps and Platform Engineering" \
--generation-system-prompt "You are an expert in DevOps and Platform Engineering" \
--mode graph \
--depth 2 \
--degree 2 \
--provider openai \
--model gpt-4o \
--num-samples 2 \
--batch-size 1 \
--conversation-type chain_of_thought \
--reasoning-style freetext \
--output-save-as dataset.jsonl \
What DeepFabric Does¶
- Generates topic hierarchies from a root prompt
- Creates training samples for each topic
- Outputs JSONL compatible with HuggingFace and training frameworks
Dataset Types¶
| Type | Description | Use Case |
|---|---|---|
| Basic | Simple Q&A pairs | Instruction following |
| Reasoning | Chain-of-thought traces | Step-by-step problem solving |
| Agent | Tool-calling with real execution | Building agents |
Key Features¶
Topic-driven generation ensures diverse, non-redundant samples. Each training example maps to a specific subtopic, avoiding the repetition common in naive generation.
Real tool execution via Spin. Agent datasets include actual tool results from isolated WebAssembly sandboxes, not simulated outputs.
Training integration with TRL, Unsloth, and HuggingFace. Use apply_chat_template to format for any model.
Built-in evaluation for testing fine-tuned models on tool-calling tasks with metrics for accuracy and correctness.
Documentation¶
- Getting Started - Installation and first dataset
- Dataset Generation - Types and configuration
- Tools - Spin components and MCP integration
- Training - Loading datasets and formatting
- Evaluation - Testing fine-tuned models
- CLI Reference - Command documentation