| # Source and content notice |
|
|
| ## Source contributions |
|
|
| DRIVE-UGC contains selected posts and images from publicly accessible Autohome automotive-community contributions archived for a 2026 research study. The records organize these contributions into labeled events for vehicle preference ranking. |
|
|
| Copyright and other rights in third-party posts and images remain with their respective creators and other rights holders. The repository does not grant a new open-content license over these source materials. Public availability of the release and source attribution are separate from the rights governing reuse. |
|
|
| ## Released representations |
|
|
| Post titles and bodies use the research archive's sanitized text, including vehicle-identity placeholders. An additional 14 contact strings in nine posts were replaced with `[CONTACT]` for release (11 explicit mobile contact numbers, one QQ identifier, and two phone-number-formatted usernames), while frozen features preserve the encoding of the archived texts before these replacements. The original selected-image archive preserves the selected image bytes and resolves images by content hash. |
|
|
| The accompanying visual arrays were computed under the study's OCR-MASK input procedure, which targets detected vehicle-identity text. They therefore represent a different image-processing stage from the selected original image bytes. These arrays contain local region vectors for the BGFP/RRP research configuration. |
|
|
| Vehicle-identity masking controls benchmark cues. Original images may retain faces, license plates, logos, or other identifying details; event IDs and content-group IDs provide pseudonymous metadata associations. These properties should guide interpretation of the released content. |
|
|
| ## Encoder provenance |
|
|
| - Text encoder: [BAAI/bge-large-zh-v1.5](https://huggingface.co/BAAI/bge-large-zh-v1.5), producing 1,024-dimensional post features. |
| - Visual encoder: [google/siglip2-so400m-patch14-384](https://huggingface.co/google/siglip2-so400m-patch14-384), producing 1,152-dimensional regional features. |
|
|
| Encoder weights are distributed through their upstream repositories, whose model cards provide documentation and applicable terms. This release contains extracted features and data-loading materials. |
|
|
| ## Source attribution |
|
|
| Autohome: https://www.autohome.com.cn/ |
|
|
| Dataset name: DRIVE-UGC — Decision-stage Ranking with Image and Verbal Evidence from User-Generated Content. |
|
|
