Metis-RL: Reinforcement Learning Dataset for Multimodal Agent Training

Metis-RL contains approximately 5.2 thousand multimodal prompts designed for training agentic models. The dataset was created by Accio-Lab for reinforcement learning via Hierarchical Decoupled Policy Optimization and was last updated in April 2026. It focuses on cultivating meta-cognitive tool use across perception, search, and mathematical or logical reasoning tasks.

Use Cases

• Training reinforcement learning agents for multimodal perception tasks based on the described prompts.
• Developing meta-cognitive tool-use capabilities in AI models based on the described search and reasoning tasks.
• Benchmarking agent performance on mathematical and logical reasoning problems based on the dataset's scope.

Strengths

• Approximately 5.2 thousand curated multimodal prompts.
• Designed for a specific, advanced training method (Hierarchical Decoupled Policy Optimization).
• Covers three distinct task domains: perception, search, and mathematical/logical reasoning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Dataset size and file formats are unspecified.

Provenance

Source

Accio-Lab

Collection Method

Created for reinforcement learning training of the Metis-8B-RL model.

Time Range

null

Freshness

Last updated 2026-04-10 09:43:27; freshness should be verified.

Geography

null