XLand-100B: 100 Billion Transitions for In-Context Reinforcement Learning

XLand-100B provides 100 billion transitions for multi-task in-context reinforcement learning, published by dunnolab for ICLR 2025. The data covers a vast range of procedurally generated tasks within the XLand-MiniGrid environment to support meta-learning research.

Use Cases

• Training transformer architectures on the sequence of observations and rewards for in-context RL.
• Benchmarking meta-reinforcement learning performance across diverse task distributions.
• Investigating scaling behaviors in reinforcement learning using the 100 billion transition records.

Strengths

• 100 billion transitions for large-scale model training
• Peer-reviewed methodology from ICLR 2025
• Apache-2.0 license for open research use

Limitations

• Significant hardware resources required to process 100B transitions
• Domain-specific to grid-world environments which may not generalize to continuous control tasks

Provenance

Source

dunnolab, ICLR 2025

Collection Method

synthetic (procedurally generated environment)

Freshness

Last updated February 2025; corresponds with ICLR 2025 publication.