Dataset Summary

This dataset contains 2k agent trajectories generated by the Qwen2.5-32B-Instruct model using smolagents library as the agent framework.

The trajectories are collected using the "first-thought prefix" method, where each trajectory is prefixed by the model's initial reasoning steps derived from Chain-of-Thought (CoT) prompting.

For more details on the method, data format, and applications, refer to the following:

• Repository: https://github.com/Nardien/agent-distillation
• Paper: https://arxiv.org/abs/2505.17612