Synthetic agent training data from scratch — forward synthesis & back-translation
Installation • Usage • Related projects • Citation
A pipeline for synthesizing high-quality agent training data from scratch. It targets cold start and data scarcity in tool-using agent scenarios: instead of mining value from existing logs, AgentSynth generates SFT-style trajectories (user intent → reasoning + tool calls) and validates them with execution-based reject sampling.
- Synthetic over manual — Strong LLMs (e.g. GPT-4, Qwen, Claude) can produce more consistent chain-of-thought and tool-calling behavior than typical human annotators.
- Execution as ground truth — Any trajectory that fails to run correctly (syntax errors, hallucinated tools or arguments) is rejected before it enters the dataset.
- Bidirectional generation — Data is produced in two ways: forward (multi-teacher: scenario → full trajectory) and back-translation (valid tool chain → reverse-engineered user query).
-
Pipe A — Forward synthesis
Seed scenario → multiple teacher models → consensus or selection → best trajectory. -
Pipe B — Back-translation
Tool definitions → valid tool-call sequences → simulated execution → LLM generates the user query that would justify that sequence. -
Pipe C — Reject sampling
Replay each candidate trajectory in a sandbox; reject on syntax errors, unknown tools/arguments, or empty observation loops.
cd AgentSynth
pip install -e .Requires Python 3.10+. See pyproject.toml and requirements.txt for dependencies.
agentsynth --help
agentsynth run -t examples/tools.json -o out.jsonl --mode back -n 2Back-translation (tool chain → user prompt): provide a JSON file of tool definitions; the pipeline builds valid chains, back-translates each to a user query, validates, and writes accepted samples to JSONL.
Forward (scenario → trajectory): use --mode forward and --scenarios <file> (JSON array or one scenario per line).
AgentSynth/
├── src/agentsynth/
│ ├── core/ # Types and config (Pydantic models)
│ ├── teachers/ # Forward teachers and back-translator
│ ├── execution/ # Sandbox and trajectory validation
│ └── generators/ # Tool-chain builder
├── examples/ # Example tools.json for runnable demo
├── tests/
├── pyproject.toml
└── README.md
- pydantic (≥2) — schema and validation
- litellm — LLM calls (OpenAI, Anthropic, Qwen, etc.)
- tenacity — retries
- click, rich — CLI
Optional: datasets for HuggingFace dataset I/O. See pyproject.toml for dev tools (pytest, ruff, mypy).
- AdaRubrics — Adaptive dynamic rubric evaluator for agent trajectories: generates task-specific dimensions and scores runs for filtering/RLHF. Use it to score and filter AgentSynth's synthesized trajectories before training or deployment.
- AgentHER — Hindsight Experience Replay for LLM agents: relabel failed trajectories into valid training data (SFT/DPO). Complements AgentSynth when you have existing failed runs to recover instead of synthesizing from scratch.
- trajectory_tokenization — ReAct with trajectory tokenization: compresses long (Thought, Action, Observation) history so long-horizon runs fit in context. Addresses context length; AgentSynth addresses data generation.
@software{agentsynth2025,
title = {AgentSynth: Industrial-Grade Agent Data Synthesis Pipeline},
author = {Ding, Liang},
year = {2025},
url = {https://github.com/alphadl/AgentSynth},
}Apache 2.0