Loong is an open-source project focused on enabling reasoning-capable AI models to bootstrap themselves by generating and verifying synthetic data. It targets AI researchers and developers looking to scale self-improvement for LLM agents, particularly in domains requiring complex reasoning and verifiable outputs. The project provides a framework and datasets to facilitate this process, aiming to discover scaling laws for agent intelligence.

How It Works

Loong employs an agent-environment loop where a Generator creates synthetic questions and answers from seed datasets. A Verifier then assesses the correctness of these generated responses, often by executing associated rationale code. A Trainable Agent learns iteratively from these verified question-answer pairs, enabling scalable self-improvement through reinforcement learning and advanced strategies. This approach allows models to learn from their own generated data, potentially reducing reliance on expensive human-labeled datasets.

Quick Start & Requirements

• Install: Primarily through Python packages and potentially Docker. Specific commands are detailed in the project's cookbooks.
• Prerequisites: Python environment, with specific dependencies varying by domain (e.g., numpy, pandas, torch). Some datasets may require specific libraries for rationale execution.
• Resources: Setup time and resource requirements depend on the scale of data generation and training.
• Links: Cookbooks, Datasets, Loong Blog.

Highlighted Details

• Includes 8,729 questions across 12 diverse domains, including Advanced Math, Physics, Chemistry, Finance, and Programming.
• Datasets are structured with questions, answers, rationales (often code), and metadata, allowing for automatic verification.
• Provides modular "Cookbooks" for synthetic data generation, verification, and RL training loops.
• Actively seeks community contributions for seed datasets, verifiers, and cookbook improvements.

Maintenance & Community

• Led by the CAMEL team, with contributions from the open-source AI research community.
• Community channels include Discord and WeChat.
• Actively seeking participants for an "Initiative Program."

Licensing & Compatibility

• Code License: LICENSE file (likely MIT or Apache, common for CAMEL projects).
• Data License: Per-dataset license specified in metadata.json.
• Compatibility: Suitable for research and potentially commercial use, depending on individual data licenses.

Limitations & Caveats

The effectiveness of the self-bootstrapping process is contingent on the quality and coverage of the initial seed datasets and the accuracy of the verifiers. Some domains may require specific computational environments for rationale execution.