CAMEL-7B Multi-Agent Instruction Tuning Dataset

Based on an improved CAMEL framework, the data was generated by introducing a behavioral constraint mechanism, a predefined library of specialized roles (e.g., Product Managers, Engineers), and structured interaction protocols to produce task-oriented multi-turn collaborative dialogues.

Grounded on 1,000 groups of high-quality multi-turn dialogues, the dataset was refined through cleaning and structured extraction to form 2,315 standardized instruction-response pairs. The data covers four main domains: code generation (45%), mathematical reasoning (30%), scientific research tasks (15%), and other comprehensive fields (10%). Experiments demonstrate that applying supervised fine-tuning (SFT) with this dataset significantly improves the performance of 1.3B to 7B scale open-source models on benchmarks such as HumanEval, exhibiting strong cross-model generalization capabilities.