AFM-Datasets: An End-to-End Chain-of-Agents Dataset
Introduction
| Dataset | AFM-Datasets |
| Modalities | Text |
| Formats | json |
| Languages | English |
| Size | 41.6 kB |
| Release Date | 2025-08-06 |
| Domain | Agent |
| License | apache-2.0 |
AFM-Datasets is the official training dataset released with the research paper, "Chain-of-Agents," and is specifically designed for building Agent Foundation Models (AFMs).
The core objective of this dataset is to train a single large language model to simulate a "multi-agent team," enabling it to solve complex tasks—such as web navigation and code generation—autonomously and end-to-end.
It primarily consists of two types of data:
- Supervised Fine-Tuning (SFT) Data: Generated through "multi-agent distillation," this data captures the complete problem-solving trajectories of state-of-the-art multi-agent systems.
- Reinforcement Learning (RL) Data: Used for agentic reinforcement learning to further enhance the model's decision-making and execution abilities on verifiable tasks.