VidHalLoc

VidHalLoc contains 2,000 factual/counterfactual video-language samples grounded in 1,090 videos.

Splits

Split	Samples	Videos
Main	1,200	654
Eval	400	218
Test	400	218

The splits are strictly disjoint at the source-video level. Records contain only video_id, task, qa_pair, leaf_label, and conflict_slot. The qa_pair object contains question, answer, and counterfactual_answer.

Video paths are deterministic:

• video_qa: videos/videoqa/{video_id}.mp4
• video_captioning: videos/video_caption/{video_id}.mp4

Inference Prompt

Use the following prompt for both VideoQA and Caption samples. Replace {question} and {candidate_answer} with the corresponding fields for one candidate at a time.

You are judging whether one candidate answer is fully supported by a video and correctly answers the question.

Question: {question}
Candidate answer: {candidate_answer}

Judge this candidate independently using only the video, the question, and this candidate answer.
Do not assume that another candidate answer exists, and do not compare this answer with any alternative answer.

Return exactly one label and nothing else:
ACCEPT — the candidate answer is fully supported by the video and correctly answers the question.
REJECT — the candidate answer is contradicted by the video, unsupported, incomplete, irrelevant, or otherwise does not correctly answer the question.
If the available video evidence is insufficient to fully support the candidate answer, return REJECT.

License and source terms

The VidHalLoc annotations, benchmark metadata, and documentation are released under the Apache License 2.0. This release also includes videos derived from multiple public source datasets; those third-party media files remain subject to the licenses and terms of their respective original sources. The Apache License 2.0 does not override third-party rights or source-dataset terms.