MiroVerse v0.1: 100K-1M Full-Trajectory SFT and DPO Research Records

MiroVerse v0.1 provides between 100,000 and 1,000,000 text records for deep research and model alignment, released by miromind-ai in January 2026. The collection includes full-trajectory data for Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to support the reproduction of MiroThinker-v0.1 benchmarks. It is specifically designed for training agentic models using the MiroTrain framework.

Use Cases

• Supervised Fine-Tuning (SFT) of LLMs using the provided research trajectories
• Preference alignment using the DPO data pairs to improve model reasoning
• Benchmarking agentic performance against the MiroThinker-v0.1 baseline

Strengths

• Scale of 100,000 to 1,000,000 records
• Dual-format support with both SFT and DPO data
• Documented methodology via Arxiv 2511.11793

Limitations

• Restricted to non-commercial use under CC BY-NC 4.0
• Optimized for a specific training framework (MiroTrain) which may limit general portability

Provenance

Source

miromind-ai

Collection Method

Full-trajectory generation for deep research agents

Freshness

Last updated January 16, 2026.

Geography

United States