Sign in to view source links and access this dataset
Description
A May 1, 2025 update added a bounding box version of Web-Hybrid data, containing 757,000 datapoints after content moderation. The dataset is provided by the organization osunlp and hosted on Hugging Face. No conversation template is applied to this version, and bounding box coordinates are normalized to [0,999].
Use Cases
Train object detection models based on the provided bounding box annotations.
Benchmark model performance on a large-scale, filtered dataset of 757k datapoints.
Fine-tune vision models using normalized coordinate data without conversation templates.
Strengths
Contains 757,000 datapoints, indicating a substantial scale for model training.
Coordinates are normalized to a consistent range of [0,999], which may simplify preprocessing.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count beyond the filtered 757k datapoints is unknown, which may limit suitability assessment.
Provenance
Source
osunlp
Freshness
Last updated 2025-05-02 12:27:52.
Access requires application and approval via email contact.