MiroVerse v0.1: A Reproducible, Full-Trajectory, Ever-Growing Deep Research Dataset
- Type
- dataset
- Venue
- HuggingFace (miromind-ai / MiroMind AI)
- Year
- 2026
- Source
- huggingface
- Access
- free
- Language
- English
- Added
- 2026-07-17T20:18:03.722218+00:00
- Verified
- 2026-07-17T20:18:03.722218+00:00
Summary
MiroVerse v0.1 is a large-scale agent dataset with 147K+ samples featuring full rollout trajectories across diverse AI agent tasks including multi-hop QA, web navigation, and scientific reasoning. Every sample includes complete execution traces with 1.9B+ tokens and 602K+ tool interactions, providing comprehensive training data for tool-using and web-browsing AI agents. It aggregates and curates data from multiple sources (Voyager, MuSiQue, HotpotQA, WebWalkerQA, MegaScience, TaskCraft, etc.) and includes both SFT and DPO data, designed to reproduce MiroThinker-v0.1's benchmark performance on Qwen3. Described in arXiv:2511.11793.
Keywords
agents deep-research multi-hop-qa web-navigation tool-use trajectories sft dpo qwen3
Topics
AI Agents / Deep Research / Multi-hop QA
Research notes
- Gated dataset (requires agreeing to share contact info). Includes 11 constituent splits with different licenses. Designed to work with the MiroTrain framework. Has a free trace rollout program to help researchers train models.