PLM-Image-Auto

Dataset Card for PLM-Image Auto [\[📃 Tech Report\]](https://arxiv.org/abs/2504.13180) [\[📂 Github\]](https://github.com/facebookresearch/perception_models/) Sythetic image captions and QAs used in PLM, please refer to the paper, Section 3, for more details. The sythetic annotations covers: SA1B, Openimages, Obejct365, ArxivQA, UCSF, PDFAcc. Dataset Structure Image Captions (SA1B, Openimages, Obejct365) Data fields are : `image_id`: a `string` feature, unique identifier for the image. `image`: a `string` feature, the actual image path in the correspoding data folder. `conversations`: a `list` of `dict` feature, with the actual prompt and caption used to train PLM. A sample from SA1B Captions looks as follows: Image Captions and QA (ArxivQA) Data fields are : `uuid`: a `string` feature, unique identifier for the sample. `image`: a `string` feature, the actual image path in the correspoding data folder. For instance `2004.07107/figure_0_0.jpg` represent the paper id `2004.07107`, Page 0, and Figure 0. `conversations`: a `list` of `dict` feature, with the actual conversation used to train PLM. A sample from SA1B Captions looks as follows: Document QA (UCSF, PDFAcc) Data fields are : `uuid`: a `string` feature, unique identifier for the sample. `image`: a `string` feature, the actual image path in the correspoding data folder. For instance `pdfa-eng-train-0045/3018390_0.jpg` represent the paper shard `pdfa-eng-train-0045`, Document 3018390, and Page 0. `conversations`: a `list` of `dict` feature, with the actual conversation used to train PLM. A sample from SA1B Captions looks as follows: Data Stats | | Sample | Type | | ----------- | ----------- | ----------- | | sa1b | 9360168 | Captions | | obj365 | 3439046 | MCQAs | | openimages | 1740864 | Captions | | arxivqa | 1859680 | Captions+ QAs | | pdfacc | 12024670 | QAs | | ucsf | 5953490 | QAs | Licensing Information This data is an output from Llama 3.2, and subject to the Llama 3.2 license (https://github.com/meta-llama