Skip to content

Dataset · BreakingWeb

Web agent trajectories in a browser benchmark

BreakingWeb/breakingweb-results-v3

From 100k to 1M operational steps of top LLM agents (Claude, GPT, Gemini) in browser tasks.

What's inside

Records of web task execution trajectories in the BreakingWeb environment with paired scenarios (clean and with interventions) across seven locally deployed sites. Inside are Browser-Use logs for Claude Sonnet 4.6 and Opus 4.7, as well as BrowserGym screenshot trajectories for Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.7; the volume ranges from 100K to 1M records.

What you can do with it

  • SFT or RL fine-tuning of multimodal web agents on screenshot actions
  • Analysis of browser agent robustness to page failures and interventions
  • Comparison of navigation strategies of different LLMs in a web environment
Tasks
Reinforcement learning
Size
100K–1M rows
Downloads this month
23.3K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("BreakingWeb/breakingweb-results-v3")