CounterVid

CounterVid is the preference dataset introduced in CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models (EMNLP 2026).

It contains short synthetic videos with controlled action variations and two complementary supervision signals:

• visual preferences (input_swap): a fixed prompt and answer paired with preferred and rejected videos;
• textual preferences (output_swap): a fixed video and prompt paired with preferred and rejected answers.

Dataset composition

Split	Total	Visual preferences	Textual preferences
train	26,167	18,398	7,769
validation	2,910	1,989	921

The 26,167 training pairs constitute the approximately 26K examples reported in the paper. The validation split is a separate held-out benchmark set.

Each MP4 is stored once and may be referenced by multiple preference examples or task formats. Video filenames use content-derived SHA-256 identifiers.

Loading

from datasets import load_dataset

dataset = load_dataset("aimagelab/CounterVid")
train = dataset["train"]
validation = dataset["validation"]

The videos column contains a list of videos, and video_roles gives their ordered semantics:

• visual preference: videos = [chosen, rejected];
• textual preference: videos = [context].

Fields

• id: unique preference-example identifier
• task_type: free-form, order-list, binary-choice, or multiple-choice format
• preference_signal: visual or textual
• swap_type: original input_swap or output_swap designation
• prompt: video-language prompt
• answer: target shared by a visual-preference pair
• chosen_answer, rejected_answer: textual-preference responses
• videos: one context video or an ordered chosen/rejected pair
• video_roles: ['context'] or ['chosen', 'rejected']
• base_identifier: synthetic anchor-scene identifier
• modality: video

Creation and safety

CounterVid uses real PE Video Dataset examples only to select anchor frames and captions; the released preference videos are synthetic. The generation pipeline applies automated sexual-content and broader harmful-content screening, and flagged samples are excluded from the final dataset.

Intended use and limitations

This dataset is intended for research on video-language grounding, preference optimization, action recognition, and temporal reasoning. Synthetic videos may contain generation artifacts or residual semantic ambiguity. Models trained on this dataset may still hallucinate and should not be used as the sole basis for high-stakes decisions.

For code and complete training details, see the CounterVid repository and project page.

Citation

@inproceedings{poppi2026countervid,
  title     = {{CounterVid}: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models},
  author    = {Poppi, Tobia and Uzkent, Burak and Garg, Amanmeet and Porto, Lucas and Kessler, Garin and Yang, Yezhou and Cornia, Marcella and Baraldi, Lorenzo and Cucchiara, Rita and Schiffers, Florian},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

This work was conducted during Tobia Poppi's Applied Scientist internship with the Amazon Prime Video team in Seattle. GitHub profile.