Dataset · lukakuma
Raw web agent trajectories with images
lukakuma/web-trajectories-preference-raw
Large raw archive of successful and failed web agent actions in English and Korean for subsequent filtering.
What's inside
The dataset contains between 1 million and 10 million records: raw web navigation trajectories, texts, tables, and screenshots in English and Korean. It includes 8,851 successful and 5,960 failed trajectories, alongside automatically selected draft fine-tuning configs and unverified visual candidates. The author emphasizes this is not a ready-to-use training corpus, but a raw inventory requiring manual review and filtering.
What you can do with it
- Data selection and filtering for training multimodal web agents
- Building custom fine-tuning or preference datasets for browser navigation
- Analyzing common errors and failures during web scenarios execution
- Tasks
- Text generation
- Languages
- English, Korean
- Size
- 1M–10M rows
- Downloads this month
- 33K
- Likes
- 0
- Access
- on request on Hugging Face
How to load it
from datasets import load_dataset
ds = load_dataset("lukakuma/web-trajectories-preference-raw")