WebArbiter: Two-Stage Training Data for Process Reward Models

WebArbiter Training Data provides step-level preference pairs for training a principle-guided reasoning Process Reward Model (PRM) for web agents. The data builds on the WebPRM Collection, which comprises approximately 30,000 step-level preference pairs drawn from the Mind2Web environment. The dataset was published by ZYao720 in conjunction with an ICLR 2026 paper.

Use Cases

• Training process reward models based on step-level preference pairs.
• Evaluating web agent reasoning performance based on annotated step-level preferences.
• Benchmarking reinforcement learning algorithms for web navigation tasks.
• Developing principle-guided reasoning models for autonomous web agents.

Strengths

• Builds on a substantial base collection of approximately 30,000 step-level preference pairs.
• Associated with a peer-reviewed publication presented at ICLR 2026.
• Specifically designed for training a principle-guided reasoning model.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and exact data size are unknown, which may limit suitability assessment.
• License information is unknown, which could restrict usage.

Provenance

Source

ZYao720

Collection Method

Derived from the WebPRM Collection (Chae et al., 2025), which comprises step-level preference pairs from the Mind2Web environment.

Freshness

Last updated 2026-04-09 18:19:41; freshness should be verified.