README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - web-agent
  - process-reward-model
  - preference-data
  - SFT
size_categories:
  - 1K<n<10K

surfPRM-data

SFT training data for surfPRM, a process reward model (PRM) for web agents.

• 9,821 pairwise step-preference samples stored as a single JSON array
• Each sample has one field conversation with system / user / assistant turns: the user turn contains the task intent, the current page's accessibility tree (AXTree), the previous action trajectory, the start/current URLs, and two candidate next actions; the assistant turn contains structured <State>/<Criteria>/<Analysis>/<Answer> reasoning ending in the preferred action
• The exact prompt template is documented in the model cards of the models trained on this data: surfPRM-qwen3.5-4b and surfPRM-qwen3.5-9b

License

Released under Apache 2.0.