climate-nlp/multimodal-oil-gas-benchmark

Dataset Card for Dataset Name

This is a dataset to reproduce our work, "A Multimodal Benchmark for Framing of Oil & Gas Advertising and Potential Greenwashing Detection", accepted to NeurIPS 2025 Benchmarks and Datasets Track.

Dataset Details

Dataset Sources [optional]

• Repository: https://github.com/climate-nlp/multimodal-oil-gas-benchmark
• Paper: A Multimodal Benchmark for Framing of Oil & Gas Advertising and Potential Greenwashing Detection (NeurIPS 2025)

Uses

We created this annotated video dataset to benchmark VLMs for predicting obstruction and impressionistic framing by O&G entities.

Direct Use

Research purpose.

Out-of-Scope Use

This dataset is made available for research purposes only and may not be used for commercial purposes.

Dataset Structure

The dataset files are formatted in JSON-Line, one video sample for each line. The following is the basic data structure for each video. You might see additional fields for meta data.

{
  "video_id": "video_001",  # The unique ID for the video
  "video_url": "[ANONYMIZED]",  # The URL of the video
  "labels": ["PA"],  # The annotated label(s)
  "video_length_seconds": 15,  # The video length
  "entity_name": "[ANONYMIZED]"  # The entity name of the video publisher
}

Dataset Creation

Curation Rationale

Read the above.

Source Data

The labels are annotated for video URLs sourced from YouTube and Facebook ads.

Data Collection and Processing

Read our paper for more detail.

Who are the source data producers?

The source videos are obtained from public ad videos on the social media. The labels of the Facebook dataset are originally created in previous literature (c.f., Holder et al. and Rowlands et. al.). Read our paper for more detail.

Annotation process

The labels for the YouTube dataset are manually annotated. The labels of the Facebook dataset are distantly annotated. Read our paper for more detail.

Who are the annotators?

For the YouTube dataset, the authors engaged in the annotations. The labels of the Facebook dataset are originally annotated in previous literature (c.f., Holder et al. and Rowlands et. al.).

Personal and Sensitive Information

N/A.

Bias, Risks, and Limitations

We do not see significant risks. Read our paper for more detail.

Recommendations

N/A.

Citation

BibTeX:

@inproceedings{morio-etal-2025-multimodal,
    author = {Morio, Gaku and Rowlands, Harri and Stammbach, Dominik and Manning, Christopher D and Henderson, Peter},
    booktitle = {Advances in Neural Information Processing Systems},
    title = {A Multimodal Benchmark for Framing of Oil \& Gas Advertising and Potential Greenwashing Detection},
    year = {2025}
}

Dataset Card Contact

Please contact to the authors. The primary contact email can be found in the paper.