Hermeneia/AD4Edu-Preferences

Dataset Card for AD4Edu Preferences

Preference pairs over audio descriptions (AD) of slide-based lecture videos, for blind and low-vision (BLV) students. Each pair is two candidate ADs for the same lecture moment; the task is to say which better serves a BLV listener under the project's 45-rule lecture-AD standard (rules_for_slides.yaml; six categories: style, terminology, length, deixis, faithfulness, non-redundancy).

PRIVATE, derived from copyrighted lecture video. Do not redistribute. Keyframe images are not included here: the moment context in prompt is text (transcript window, slide OCR, an on-screen description, moment type, pause), and dpo_prompts.jsonl references frames by relative path (keyframes/<lecture>/<frame>.jpg) resolved against the separate private repo Psychias/AD4Edu-keyframes.

Lectures

The material is recorded university lectures delivered over slides, in English, across six subjects -- the same corpus as Psychias/AD4Edu-SFT. Each pair sits on one lecture moment; a moment can carry several pairs, so moments, not pairs, are the independent units. The two splits are moment-disjoint but share 16 lectures, because the validation set draws on both split sides of the corpus.

subject	train_prefs lectures	train_prefs moments	train_prefs pairs	test_human lectures	test_human moments	test_human pairs
Anatomy	19	223	1025	6	33	54
Biology	12	203	984	5	31	42
Machine learning	4	206	519	11	176	260
Mathematics	5	91	538	0	0	0
Psychology	9	85	474	5	31	44
all	49	808	3540	27	271	400

Moment types, as classified on the slide stills:

moment type	train_prefs pairs	test_human pairs
pointing	2033	286
slide	850	46
figure	288	46
reveal	183	6
animation	67	7
ink	32	4
math	36	0
table	21	1
chart	17	3
transition	13	1

Columns

train_prefs (config default):

column	what it is
source	natural_judged (two families' candidates, ordered by a video judge in both presentation orders) or constructed_counterfactual (one axis of a compliant AD perturbed deliberately)
prompt	the moment as the SFT user turn renders it text-only: Moment type, Lecturer said, On-screen text (OCR), What is on screen, Pause available
prompt_id	sha256(prompt); on a moment that is also an SFT training moment it equals the text_only config's train_gen prompt_id on Psychias/AD4Edu-SFT, so the two datasets join
chosen, rejected	[{user: prompt}, {assistant: AD text}]; the assistant turn is the bare AD, which is what train_dpo.py trains on
chosen-model, rejected-model	the family that wrote each side (table below). On constructed pairs both name the base family: the rejected text is a perturbation of the chosen one, described by axis and what_differs
chosen-rating, rejected-rating	ADEduEval mechanical compliance of each side, in [0, 1] -- see the next section before using them
chosen-components, rejected-components	the six category scores behind the rating; null = the tier abstains on that category
pair_id, moment_id, lecture, type, axis, label_source, cross_family, what_differs	provenance; pair_id is the foreign key the judge cache and the rater package hang off

test_human (config test_human) has the same columns with option_a / option_b in place of chosen / rejected, plus stratum, labels_pending and the rater columns described under Splits.

from datasets import load_dataset
train = load_dataset("Psychias/AD4Edu-Preferences", "default")["train_prefs"]
held  = load_dataset("Psychias/AD4Edu-Preferences", "test_human")["test_human"]

ADEduEval, and the rating columns

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.

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. chosen-rating / rejected-rating (and option_a-rating / option_b-rating) are the mechanical tier on each side, computed by ad4edu.validation.adedueval_eval.rate_pairs; the six components ship beside them. They are not how chosen was decided -- the direction comes from label_source (judge or construction). To recompute or extend them:

# recompute the shipped ratings from the Hub and refuse on any mismatch
python scripts/eval_adedueval.py hf --repo Psychias/AD4Edu-Preferences --config default --check-shipped
python scripts/eval_adedueval.py hf --repo Psychias/AD4Edu-Preferences --config test_human --check-shipped

# the same on a downloaded parquet, writing per-side rows and the per-axis summary
python scripts/eval_adedueval.py pairs --pairs data/train_prefs-00000-of-00001.parquet --out sides.jsonl --md summary.md

# the model-backed tier (faithfulness and non-redundancy too; local models, no API)
python scripts/eval_adedueval.py --tier local pairs --pairs data/train_prefs-00000-of-00001.parquet

The --gold context defaults to the canonical rows shipped as gold_eval.jsonl on Psychias/AD4Edu-SFT; the deixis component abstains on lectures whose recording renders no cursor (analysis/cache/cursor_capability.json).

The chosen-rating and rejected-rating columns are the mechanical tier on one description at a time. Faithfulness and non-redundancy are null because they need the model-backed tier, and within-lecture novelty does not apply, because the two sides of a pair come from different systems and a rating here is not a system score. So these columns penalise defects and oblige nothing: do not filter or reweight pairs by rating margin and expect a quality signal. What the rating does tell you, per axis, is how often the compliant side of a controlled pair also scores higher -- measured at build time on this split, strata before any aggregate:

by axis (strata first)

stratum	pairs	moments	directed	ties	unscored	agree on decided	agree, ties = fail
deixis	219	114	219	25	0	82.5% [76.5, 87.2] (k=160, n=194)	73.1% [66.8, 78.5] (k=160, n=219)
faithfulness	145	93	145	28	0	90.6% [83.9, 94.7] (k=106, n=117)	73.1% [65.4, 79.7] (k=106, n=145)
length	191	117	191	2	0	88.9% [83.6, 92.6] (k=168, n=189)	88.0% [82.6, 91.8] (k=168, n=191)
natural	2501	759	2501	58	0	52.4% [50.5, 54.4] (k=1281, n=2443)	51.2% [49.3, 53.2] (k=1281, n=2501)
non_redundancy	170	106	170	17	0	43.8% [36.2, 51.7] (k=67, n=153)	39.4% [32.4, 46.9] (k=67, n=170)
style	149	92	149	4	0	82.8% [75.8, 88.0] (k=120, n=145)	80.5% [73.4, 86.1] (k=120, n=149)
terminology	165	99	165	100	0	50.8% [38.9, 62.5] (k=33, n=65)	20.0% [14.6, 26.8] (k=33, n=165)

by type (strata first)

stratum	pairs	moments	directed	ties	unscored	agree on decided	agree, ties = fail
animation	67	19	67	5	0	62.9% [50.5, 73.8] (k=39, n=62)	58.2% [46.3, 69.3] (k=39, n=67)
chart	17	4	17	0	0	47.1% [26.2, 69.0] (k=8, n=17)	47.1% [26.2, 69.0] (k=8, n=17)
figure	288	63	288	22	0	69.9% [64.2, 75.1] (k=186, n=266)	64.6% [58.9, 69.9] (k=186, n=288)
ink	32	6	32	3	0	41.4% [25.5, 59.3] (k=12, n=29)	37.5% [22.9, 54.7] (k=12, n=32)
math	36	15	36	0	0	44.4% [29.5, 60.4] (k=16, n=36)	44.4% [29.5, 60.4] (k=16, n=36)
pointing	2033	418	2033	165	0	59.8% [57.6, 62.0] (k=1117, n=1868)	54.9% [52.8, 57.1] (k=1117, n=2033)
reveal	183	62	183	3	0	40.6% [33.7, 47.9] (k=73, n=180)	39.9% [33.1, 47.1] (k=73, n=183)
slide	850	221	850	35	0	56.3% [52.9, 59.7] (k=459, n=815)	54.0% [50.6, 57.3] (k=459, n=850)
table	21	7	21	1	0	75.0% [53.1, 88.8] (k=15, n=20)	71.4% [50.0, 86.2] (k=15, n=21)
transition	13	2	13	0	0	76.9% [49.7, 91.8] (k=10, n=13)	76.9% [49.7, 91.8] (k=10, n=13)

Aggregate, pooling axes ['deixis', 'faithfulness', 'length', 'natural', 'non_redundancy', 'style', 'terminology'] and types ['animation', 'chart', 'figure', 'ink', 'math', 'pointing', 'reveal', 'slide', 'table', 'transition']:

stratum	pairs	moments	directed	ties	unscored	agree on decided	agree, ties = fail
ALL	3540	808	3540	234	0	58.5% [56.8, 60.2] (k=1935, n=3306)	54.7% [53.0, 56.3] (k=1935, n=3540)

Label provenance -- read before using train_prefs

label_source	n	how the direction is known
construction	1039	one axis perturbed deliberately; direction true by construction, no judge involved
judge	2501	Gemini judged a 16 s clip in both presentation orders and agreed with itself

The judge was measured before use (pooled over both judge runs, n = 2,961 pairs; analysis/j1_results.md): 14.7 % of verdicts reversed under order swap [13.5, 16.0] and are excluded, which is why both orders are mandatory; the longer side wins 56.9 % [54.6, 59.2] of cross-family decided pairs (+6.9 pp), small but real, and it will bias any length-sensitive comparison these labels feed; the judge's own family wins 54.9 % against Claude's 67.0 %, so self-preference is not borne out. Pairs the judge could not resolve (flips, ties, unparsed) are not here. These caveats apply to the 2501 judge-labelled pairs; the 1039 construction-backed pairs carry neither. 53 natural pairs are same-family (cross_family: false).

Model columns carry the repo's family key (the unit cross_family is defined on and the suffix of every output_id); the provider id behind each key:

family key	provider	model id
claude-sonnet-4.6	openrouter	anthropic/claude-sonnet-4.6
gemini-3.1-pro	gemini	gemini-3.1-pro-preview
gemini-3.1-pro-or	openrouter	google/gemini-3.1-pro-preview
gpt-5.5	openrouter	openai/gpt-5.5
qwen3-vl-235b	openrouter	qwen/qwen3-vl-235b-a22b-instruct

Splits

config / split	pairs	moments	labelled?	what it is
default / train_prefs	3540	808	yes (machine)	binarised chosen / rejected, for DPO; the train_eligible rows of dpo_train.jsonl
test_human / test_human	400	271	yes (four BLV raters)	human preference labels; the evaluation set; strata {'natural_flip': 50, 'controlled': 250, 'natural_decided': 100}

test_human carries the four blind/low-vision raters' choices as rater_1..rater_4 (the side each chose, in the row's own option_a / option_b terms, or tie) and human_winner: the side chosen by the majority of the raters who decided (a tie does not count as a vote), null when the deciding raters split evenly. Per-rater columns ship regardless, so nothing is hidden behind the aggregate. labels_pending is false on every row. Two machine annotators' choices ship beside them as machine_<family> (side or tie) with machine_<family>_reason; they are evaluation columns and never the label. These are the labels the metric-validation tables are computed against; the split never enters training.

The two splits are moment-disjoint, not merely pair-disjoint: a moment's clip, transcript window and references are shared across all its pairs, so effective n is moments. 0 eligible rows had no direction and were dropped.

Course-level holdout: MATHEMATICS, DENTAL and MACHINE LEARNING are held out whole, derived from the course map (validation.split.is_c1_test), with the reviewed lecture-level exceptions in TRAIN_OVERRIDE; the validation set draws on both sides.

Raw files

Beside the parquet: dpo_train.jsonl (every candidate pair, with the flags that make train_eligible inspectable), dpo_prompts.jsonl (the prompt of every moment these pairs and the validation set reference, in the SFT chat format with keyframe paths -- what on-policy sampling and preference evaluation join on), human_val_400.jsonl.

Citation