Hermeneia/AD4Edu-keyframes

Dataset Card for AD4Edu keyframes

The slide keyframes behind Psychias/AD4Edu-SFT and Psychias/AD4Edu-Preferences: one JPEG per sampled instant of a lecture recording, laid out as <lecture>/<frame>.jpg where the frame name is the wall-clock position in the video as HHMMSS. A moment's rows in the two datasets reference up to three frames taken in the window from two seconds before to five seconds after the moment -- the slide as it looks right after the change the moment describes.

PRIVATE, derived from copyrighted lecture video. Do not redistribute. The frames are stills of the recordings; the recordings themselves are never redistributed here.

Lectures

The material is recorded university lectures delivered over slides, in English. Frames are stored for every lecture the corpus was built from, whether or not the released datasets reference them; the last column counts the frames the released rows point at. Rows with a zero there are sources the corpus kept but the released datasets do not draw on (short conference talks rather than lectures).

subject	lecture folders	frames	frames referenced by the released datasets
Anatomy	19	3580	1109
Biology	12	2174	1064
Dental	4	692	160
Human computer interaction	1	35	0
Machine learning	15	3006	1003
Mathematics	10	1085	216
Psychology	9	1345	316
Ubiquitous computing	4	241	0
all	74	12158	3868

Frames referenced per released file:

file	frames referenced
dpo_prompts.jsonl	2055
sft_eval.jsonl	674
sft_prompts.jsonl	3157
sft_train.jsonl	3157

Every referenced frame is present in this repo.

How the paths join

Rows carry the path relative to a media root: keyframes/<lecture>/<frame>.jpg in the images column of Psychias/AD4Edu-SFT (and in the image parts of the raw sft_*.jsonl, sft_prompts.jsonl and dpo_prompts.jsonl). Place this repo at <media root>/keyframes/ -- for the project's own scripts that is $AD4EDU_DATA_DIR/preprocessing/keyframes/, where scripts/fetch_data.sh puts it -- and every path resolves. scripts/check_keyframes.py verifies that in seconds before a training run.

hf download Psychias/AD4Edu-keyframes --repo-type dataset --local-dir "$AD4EDU_DATA_DIR/preprocessing/keyframes"
python scripts/check_keyframes.py     # "OK: all keyframe references resolve."

Frames are added only, never deleted, by scripts/push_keyframes.py, which uploads the lecture folders the released rows reference and this repo lacks.

ADEduEval

ADEduEval is a diagnostic benchmark for evaluating audio descriptions of lecture content. It operationalizes a 45-rule rubric across six dimensions—style, terminology, length, deixis, faithfulness, and non-redundancy—and returns aggregate, category-level, and rule-level results. Rather than providing only a single quality score, ADEduEval identifies the specific criteria that each candidate description satisfies or violates. Its three evaluation tiers are available through a single entry point: scripts/eval_adedueval.py (Python module: ad4edu.validation.adedueval_eval).

Licensing

The synthetic ADs and the rule standard used with these frames are released under CC BY-NC-SA 4.0, consistent with both source datasets — LPM and M3AV are each CC BY-NC-SA 4.0 (LPM's videos are additionally under the standard YouTube License). The keyframes in this repository are source-derived lecture stills and are not covered by the license for the synthetic ADs or rule standard. Source videos and full per-lecture transcripts/OCR are not redistributed. Confirm per-source terms before any public release and prefer reconstructing source-derived material from the original corpora over redistributing it.

Citation

Cite AD4Edu and the two source datasets it is derived from.

@misc{ad4edu2026,
  title  = {AD4Edu: Benchmarking Audio Description Quality for Slide-Based Lecture Videos},
  author = {Psychias, Stelios},
  year   = {2026}
}

@inproceedings{lee2023lpm,
  title     = {Lecture Presentations Multimodal Dataset: Towards Understanding Multimodality in Educational Videos},
  author    = {Lee, Dong Won and Ahuja, Chaitanya and Liang, Paul Pu and Natu, Sanika and Morency, Louis-Philippe},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2023},
  note      = {arXiv:2208.08080}
}

@inproceedings{chen2024m3av,
  title     = {{M3AV}: A Multimodal, Multigenre, and Multipurpose Audio-Visual Academic Lecture Dataset},
  author    = {Chen, Zhe and Liu, Heyang and Yu, Wenyi and Sun, Guangzhi and Liu, Hongcheng and Wu, Ji and Zhang, Chao and Wang, Yu and Wang, Yanfeng},
  booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages     = {9041--9060},
  year      = {2024},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/2024.acl-long.489},
  note      = {arXiv:2403.14168}
}