Hermeneia/AD4Edu-SFT

Dataset Card for AD4Edu SFT

Instruction-tuning data for generating audio descriptions (AD) of slide-based lecture videos for blind and low-vision students. Each example is one lecture moment -- a slide change, a pointing gesture, a reveal -- with the visual context, the lecturer's speech around it, and the available pause. The target is a JSON object: whether to describe, and if so the description.

PRIVATE, derived from copyrighted lecture video. Keyframe images are not in this repo: the images column carries relative paths (keyframes/<lecture>/<frame>.jpg) resolved against a separate private repo, Psychias/AD4Edu-keyframes.

Lectures

The material is recorded university lectures delivered over slides, in English, across six subjects. Each lecture was segmented into moments where the visual content changes; every moment carries one target per teacher family that answered for it, so a moment contributes up to three rows.

subject	training lectures	training moments	training rows	validation-only lectures	validation-only moments	validation-only rows
Anatomy	19	840	1320	0	0	0
Biology	12	734	1799	0	0	0
Dental	0	0	0	4	91	273
Machine learning	4	399	1197	11	435	1255
Mathematics	5	119	119	2	55	165
Psychology	9	239	528	0	0	0
all	49	2331	4963	17	581	1693

Moment types, as classified on the slide stills:

moment type	training rows	validation-only rows
pointing	2204	1230
slide	1774	207
reveal	485	45
figure	242	76
math	90	78
ink	30	57
animation	64	0
transition	30	0
table	22	0
chart	19	0
other	3	0

Columns

column	what it is
prompt	the user turn's text: Moment type, Lecturer said (the transcript window), On-screen text (OCR), What is on screen (text_only only), Pause available
prompt_id	sha256(prompt)
messages	[system, user, assistant]; the system turn carries the operative standard the model is trained to follow, the assistant turn the JSON target. train_gen rows stop at the user turn
images	keyframe paths of the user turn, in order (empty in text_only)
output_id, segment_id, moment_id, lecture	provenance; output_id is moment::family
type	the moment type (table above)
family	which teacher family wrote this row's target
coverage, n_families_attempted, n_families_emitted	how much of the moment the speech already covers, and how many teacher families were asked for / answered on this moment
emit, rung	parsed from the target: whether the moment is described, and the delivery rung

prompt_id identifies a prompt text. A moment's rows can differ in their transcript window (the references were generated under different evidence windows), so one moment can carry several prompt_ids; conversely two adjacent moments with the same transcript window, OCR, type and pause render the same text (in default their images differ). train_sft has 4963 rows over 2331 moments and 2974 distinct prompts; 19 prompt texts are shared by 36 moments. Join on moment_id when the moment is what you mean.

from datasets import load_dataset
ds = load_dataset("Psychias/AD4Edu-SFT")                 # multimodal: train_sft / test_sft / train_gen
ds = load_dataset("Psychias/AD4Edu-SFT", "text_only")    # no images
ds = load_dataset("Psychias/AD4Edu-SFT", "cot")          # rationale before the JSON

Splits

config	train_sft	test_sft	train_gen	inputs
default	4963	1693	2331	keyframes + OCR + transcript + type + pause
text_only	4963	1693	2331	as above, images replaced by an on-screen description
cot	4963	1693	2331	as default, rationale emitted before the JSON

All three configs cover identical output_ids (asserted at build). train_gen is prompt-only: one row per train_sft moment with the assistant turn removed, for sampling or scoring a model on the training prompts without the targets.

Some lectures are used only for validation. The split is by course, not by lecture. Dental (4 lectures) is held out entirely: its lectures appear in test_sft only and are never trained on. Machine learning and Mathematics are held-out subjects as well, except for a reviewed set of lectures moved to the training side (Machine learning: 4 training, 11 validation; Mathematics: 5 training, 2 validation); a lecture is on one side only. train_sft spans 49 lectures and test_sft 17; no lecture is in both, so a model evaluated on test_sft has seen none of those lectures.

The target

One head, four keys:

{"emit": true, "ad_text": "...", "rung": 3, "rationale": "..."}

emit=false with ad_text=null means the moment should not be described -- the lecturer's speech already covers it. Suppressions are supervised, not filtered out (3062 of 4963 training rows), and each carries the reason in rationale. emit/suppress is an intermediate step; the benchmark scores AD text under forced generation. In cot the rationale precedes the JSON.

ADEduEval

ADEduEval is a metric: it grades an audio description against the 45-rule lecture-AD standard along six categories -- style, terminology, length, deixis, faithfulness, non-redundancy -- and reports where it breaks a rule. It has three tiers, all reachable through one script, scripts/eval_adedueval.py (module ad4edu.validation.adedueval_eval):

tier	what it computes	cost
mech (default)	style, terminology, length, deixis from rules alone; the other two abstain	none, deterministic
local	all six, adding a local NLI model for faithfulness and a sentence encoder for non-redundancy	none, local models
rubric	an LLM judge grading the six categories against the standard and the reference descriptions	paid; refused without --i-approve-spend

A seventh term sits outside those six and is part of the same score. Within-lecture novelty is one minus a description's largest overlap with what the same system already said about earlier moments in the same lecture, and it scales the mean of the applicable categories by min(1, novelty / 0.4). It is a property of a system's output sequence rather than of one description, so predictions carries it, while pairs and hf score single descriptions and the two sides of a pair come from different systems. The deixis component abstains on lectures whose recording renders no cursor, so a category that cannot apply to a moment is left out of the mean rather than scored as zero.

Every summary the script writes lists the strata (moment type, or preference axis) before any aggregate, with the number of units, coverage, and a 95 % Wilson interval. The two free tiers measure compliance, which is not quality: they say which rule a description breaks, and a description can break none while telling a blind listener nothing. A system is scored over its whole sequence of descriptions, so one that reads the slide text aloud at every moment is scored on the repetition as well as on each line, and ranks last.

On this dataset. Generate one description per test_sft moment (the moment is moment_id; a moment's reference rows share the prompt, so generate once per moment, not per row) into a jsonl with output_id, moment_id, emit and ad_text, then score it against the contexts in gold_eval.jsonl:

python scripts/eval_adedueval.py predictions --pred predictions.jsonl --gold gold_eval.jsonl --method sft --novelty-knee 0.4 --md report.md
python scripts/eval_adedueval.py --tier local predictions --pred predictions.jsonl --gold gold_eval.jsonl --method sft --novelty-knee 0.4

--novelty-knee 0.4 is the calibrated setting and the figure the project reports: the compliance mean over the applicable categories, scaled by within-lecture novelty. --gold also supplies each moment's time, since novelty needs the lecture's moments in order.

--method names how the system was trained (a run_meta.json beside the file wins over the flag); the metric refuses reward-optimised arms, whose objective it is. The report lists every moment type before any aggregate, with coverage (unfilled and missing moments are counted, never dropped) and a 95 % interval on each compliance rate. The reference targets themselves can be scored the same way by writing one row per moment from test_sft -- useful as the ceiling the metric assigns to the teachers.

Raw files

sft_{train,eval}{,_textonly,_cot}.jsonl (the chat examples with typed user parts, as the trainers read them), sft_prompts.jsonl (the prompt rows train_gen is built from) and gold_eval.jsonl (the canonical rows every context here comes from) ship beside the parquet under those names.

Citation