OpenWebRL RL Tasks contains the browser-task prompts used to train OpenWebRL agents with online multi-turn reinforcement learning on live websites. The dataset includes task specifications for training visual web agents by running browser rollouts, collecting screenshots and environment feedback, and optimizing with MM-GRPO. It was authored by OpenWebRL and last updated on June 2, 2026.
Use Cases
- Training reinforcement learning agents based on browser-task prompts
- Benchmarking web agent performance based on multi-turn task specifications
- Developing reward models for web interaction based on format and VLM-as-a-judge rewards mentioned in the description
Strengths
- Task specifications are designed for training with online multi-turn reinforcement learning on live websites.
- The dataset is associated with a specific training methodology (MM-GRPO) using format and VLM-as-a-judge rewards.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- OpenWebRL
- Collection Method
- Likely gathered from the creation of prompts for training web agents.
- Freshness
- Last updated 2026-06-02 05:32:05; freshness should be verified.