AD4Edu

Audio description (AD) of slide-based lecture videos, for blind and low-vision (BLV) students. A recorded lecture gives a BLV student the lecturer's speech but not the slides, so "as you can see here" has no referent. AD4Edu divides a lecture into moments -- a slide change, a pointing gesture, a reveal -- and asks, for each one, whether a description is needed and what it should say, under a 45-rule lecture-AD standard (rules_for_slides.yaml; six categories: style, terminology, length, deixis, faithfulness, non-redundancy).

This repository holds no data. It lists what AD4Edu releases and where each part lives, so that one address reaches all of them.

The datasets derive from copyrighted lecture video and are for research use only. Do not redistribute them. No recording is redistributed anywhere: the datasets carry text, slide stills, and the video identifiers and timestamps that locate each moment. The code, the two tools and the adapters are under their own licences, listed at the end.

Datasets

repository	contents	size
Hermeneia/AD4Edu-SFT	the reference descriptions and the training examples built from them. One row per moment and teacher family: the slide OCR, the transcript window, the available pause, up to three keyframe paths, and a JSON target that says whether to describe and, if so, what to say. Three configs: default (with keyframes), text_only, cot	train_sft 4963 rows, test_sft 1693, train_gen 2331
Hermeneia/AD4Edu-Preferences	preference pairs, two candidate descriptions of one moment. Natural pairs are ordered by a video judge that saw both presentation orders; controlled pairs differ on one rule category by construction. test_human is the 400 pairs rated by four BLV raters, with every rater's choice	train_prefs 3540 pairs, test_human 400
Hermeneia/AD4Edu-keyframes	the slide stills the two datasets reference by path: one JPEG per sampled instant of a recording, laid out as <lecture>/<frame>.jpg	12158 frames in 74 lecture folders, of which 3868 are referenced by the released rows

The sizes are the split counts in each card's front matter.

Models

Five LoRA adapters for Qwen3-VL, each in a seed0/ folder of its repository. The text-only arm reads the slide text and the transcript; the multimodal arm also reads the keyframes. A DPO adapter was trained on the merged weights of the supervised adapter of its row, so it is loaded on top of that adapter and not on the bare base model; each model card shows the order.

adapter	backbone	input arm	method
Hermeneia/ad4edu-qwen3vl-2b-sft	Qwen3-VL 2B Instruct	multimodal	supervised fine-tuning
Hermeneia/ad4edu-qwen3vl-8b-sft	Qwen3-VL 8B Instruct	multimodal	supervised fine-tuning
Hermeneia/ad4edu-qwen3vl-2b-r3-dpo	Qwen3-VL 2B Instruct	multimodal	direct preference optimisation, recipe r3
Hermeneia/ad4edu-qwen3vl-2b-textonly-r3-dpo	Qwen3-VL 2B Instruct	text-only	direct preference optimisation, recipe r3
Hermeneia/ad4edu-qwen3vl-8b-textonly-r3-dpo	Qwen3-VL 8B Instruct	text-only	direct preference optimisation, recipe r3

r3 is the name of the DPO recipe, as recorded in each adapter's seed0/run_meta.json: learning rate 1.5e-5 for 2 epochs at beta 0.1 over 2742 training pairs, starting from a supervised adapter trained at 1e-4 for 3 epochs. The 2B multimodal supervised adapter is the one the lecture player ships.

Code and tools

repository	contents
psychias/ad_for_edu	the code base: corpus construction from recordings (transcript, pauses, slide text, keyframes), moment detection and classification, description generation, pair building and judging, the source file of the 45-rule standard, the AD4Edu-Eval metric, training, and the analyses of the paper. One command-line entry point; every command that would call a hosted model prints its cost and exits until the spend is approved
psychias/blv_annotation_tool	the rating tool of the BLV study: one HTML file that runs offline in the browser, with the two descriptions of each item synthesised in advance and embedded. Built for screen readers from the start, WCAG 2.1 AA, every action on a single key. It takes a sheet of paired texts and returns a small answers file, and a second script reads the answers back out
psychias/blv_mediaplayer	the lecture player: it prepares a lecture once, speaks each description into a gap in the lecturer's speech, and writes WebVTT captions for the lecturer and for the descriptions. Runs the 2B multimodal adapter offline on macOS with Apple Silicon; keyboard-operable and works with VoiceOver

How the parts join

• A moment is a moment_id, the same string in both datasets. A moment can carry several rows and several pairs, so moments, not rows or pairs, are the independent units.
• prompt_id is sha256(prompt). On a moment that is also a training moment it is the same in AD4Edu-Preferences and in the text_only config of AD4Edu-SFT, so the two datasets join on it.
• Keyframes are referenced as keyframes/<lecture>/<frame>.jpg, relative to a media root. Place AD4Edu-keyframes at <media root>/keyframes/ and every path resolves.
• The code reads the dataset ids from AD_FOR_EDU_REFERENCES_DATASET, AD_FOR_EDU_PREFERENCES_DATASET and AD_FOR_EDU_KEYFRAMES_DATASET. Set them to the three repositories above.

from datasets import load_dataset
sft   = load_dataset("Hermeneia/AD4Edu-SFT")                                      # train_sft / test_sft / train_gen
prefs = load_dataset("Hermeneia/AD4Edu-Preferences", "default")["train_prefs"]
human = load_dataset("Hermeneia/AD4Edu-Preferences", "test_human")["test_human"]

hf download Hermeneia/AD4Edu-keyframes --repo-type dataset --local-dir "<media root>/keyframes"

Licences

resource	licence
AD4Edu-SFT, AD4Edu-Preferences, AD4Edu-keyframes	research use only; derived from copyrighted lecture video and not to be redistributed
ad_for_edu, blv_annotation_tool, blv_mediaplayer	Apache-2.0
the five adapters	Apache-2.0, the same as the base model

Citation

A paper describing the benchmark is under review. Until it appears, cite the project by its address:

@software{ad4edu,
  title = {AD4Edu: audio description for slide-based lecture recordings},
  url   = {https://github.com/psychias/ad_for_edu},
  year  = {2026}
}