This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.

D22-goods-annotated

The explain view of AI4Manufacturing/D22-goods: the same 3,115 train-split good records of PKU-GoodsAD (the only D22 records that may enter a training pool; every anomaly is designated test and eval-locked), with a short, gated, hedged-absence explain prose purchased for each good (ruling 2026-09-08: goods carry BOTH the terse answer-only view and an explain view; the reason is register coverage). query, image, annot, mask, cate, task are byte-identical to D22-goods; row order, shard names and row-group sizes are preserved.

Records

category	records	with explain prose	answer-only	judge FLAG	judge unjudged	defect kinds the prose rules out (display phrases)
`cigarette_box`	183	183	0 (0.0 %)	0	0	opened
`drink_bottle`	719	685	34 (4.7 %)	33	1	cap half open, cap open, surface damage
`drink_can`	235	231	4 (1.7 %)	4	0	deformation, straw missing, surface damage
`food_bottle`	1,006	948	58 (5.8 %)	51	7	deformation, opened, surface damage
`food_box`	432	425	7 (1.6 %)	7	0	deformation, opened, surface damage
`food_package`	540	540	0 (0.0 %)	0	0	broken, surface anomaly
total	3,115	3,012 (96.7 %)	103 (3.3 %)	95	8

Answer-only = 95 records the judge flagged + 8 records the judge left unjudged (its reply was truncated at 200 tokens while deliberating; an instrument failure, not a verdict on the text). Both route answer-only under the user's standing rule (with sufficient data, the strict option on suspect samples); the record keeps its gold and image, reasoning is null, and metadata.explain.reason says which of the two it was. No re-judge of the unjudged rows and no second judge vote were run (offered as options; not taken).

Roles

Roles: two answer VIEWS render from ONE purchased prose (design record §3 element 3, §7 step 7). Decomposed view — the card convention: where reasoning is present it is the SFT imitation target and its FINAL ANSWER segment (FINAL ANSWER: {good, null}, the form the query requests) is the model-facing format; annot is the machine-parseable gold. Unified view — reasoning_unified is the imitation target whose FINAL ANSWER line is good (the unified grammar's good form), requested by query_unified; annot_unified (good) is its machine-parseable gold. Neither annot column is an output-format target. On the 103 answer-only rows reasoning and reasoning_unified are null: those rows train annot / annot_unified directly in the query-specified format (the answer-only tier); a consumer selecting the explain register filters reasoning IS NOT NULL. The teacher never wrote a FINAL line — both are rendered at assembly from the record's own gold (annotate/common/render_unified.py), so the prose body is byte-identical in the two columns.

How the prose was made

Teacher: gpt-5.6-terra @ medium reasoning, official OpenAI Batch API, one full-frame image at ≤ 1024 px, prompt sha16 df0b9614a5f6e7e8 (annotate/D22/annotate_goods.py::SYSTEM/USER). Gold-conditioned: the teacher is told the item is good and which defect kinds its category is checked for, and writes 2–4 sentences of concrete first-person observation plus ONE hedged closing sentence naming the listed kinds ("no X, Y or Z of the listed kinds is apparent in this view"), with a NOT-VISIBLE hatch for parts the photo does not show. The hedge lives only in the closing sentence; unhedged completeness claims are forbidden. The teacher never sees the published query.
Display names: the teacher was shown the category's defect kinds as display phrases — the de-underscored source tokens (surface_damage → "surface damage", straw_missing → "straw missing" …; PKU-GoodsAD's own wording IS the token, so nothing is invented and the map is invertible). Every record carries metadata.explain.display_table (source, sha16 c87607da371f6d17, the category's phrases). Those phrases were checked against the corpus-wide table common/display_names.py (commit 62cb8a722996, family "GoodsAD (D22)", repo id D22) at publish time: its five non-natural rows agree with this family's phrases byte for byte, and the three natural words (opened, broken, deformation) carry no row there by design — identity is the answer for them (0 disagreements). Ruling E5 (model-facing prose carries the phrase) is satisfied by construction here rather than by substitution: the teacher never saw the raw query or a source token, so the purchased bodies contain 0 underscored tokens (measured over all 3,012 rows) and there is no pre-substitution body to store. Ruling E1-a — the published query still lists the underscored source tokens, carried byte-identically from D22-goods; converting queries is the corpus-wide requery wave's job, on this repo as on every other, and annot keeps the source form regardless.
Deterministic pre-filter ($0): leak/scaffold echoes, underscored tokens, unhedged completeness claims, global rules, an unhedged closing, list formatting, length. Self-tested on 11 planted faults. Its rules evolved on this run: a first pass (forge_model commit c19c130 (annotate/D22/annotate_goods.py::prefilter_one)) flagged 20 records; every flag was read and three rules were found to match form rather than meaning (16 "drink can" echoes were all metal cans; 3 "gold label" hits were a colour; "fully intact" described a part) and were refined with both-direction proofs (forge_model commit 3e80bb3 (prefilter_one + prefilter_notes)); final: 0 flags, 16 notes handed to the judge (which passed 15 of the 16 with the image).
Judge (a different VENDOR from the teacher, ruled 2026-09-18): claude-sonnet-5 on the official Anthropic Batch API, thinking off, rubric sha16 ea39593dba28f6ee (annotate_goods.py::JUDGE_GOODS), the same 1024-px render, gold-conditioned; single vote; calibrated on the 24-record smoke under this rubric (24 PASS). Verdicts: PASS 3,012 · FLAG 95 · unjudged 8.
Known judge weakness, disclosed: every one of the 117 flagged/unjudged records was read against its image by the annotating session (13 labelled contact sheets + 4 zooms; a weaker instrument than the judge's, so class-level reads, not per-record overrides). No gold noise was found (nothing is actually opened or damaged). drink_bottle's flags are a stock-phrase over-assertion ("with the tamper band in place beneath it") — the judge is right in spirit. food_bottle's flags are in the majority the judge misreading the folded pull-tab of foil lids (Oreo mini, Nissin, 汤达人 cups, often under shrink-wrap) as "peeled", with a minority of genuine over-assertions (a lid described when the top is not in view); the fair prose lost there is the price of the strict rule — no data is lost. Reads: outputs/d22_goods_explain/evidence/full_flags/worker_reads_full.json (quoted in the QA page). (2026-09-22) Fourteen of those 117 have since been restored to the explain route — their first judge reply was truncated at the token cap (PARSE_FAIL, an instrument failure, not a verdict) and re-judging returned PASS. The reading described above covered all 117 as they stood then; 103 remain answer-only.
(2026-09-22) Two explain rows carry prose with one sentence corrected by hand. A human annotator reading the purchased description against the photograph found one wrong assertion in each — a cap called pale green that is pale yellow (the label is the green part), and a box called upright that is lying on its side. The machine gate had passed both. Only the wrong sentence was replaced; the rest of each chain, the gold answer, the image and the route are untouched, and each of the two rows carries metadata.reasoning_provenance naming what was replaced, with what, why, and on whose decision. No other row's metadata changes. The human check that found them did NOT pass its own control gate (1 of 8 planted controls missed = 12.5 %, over the pre-registered 10 %); it was accepted by an explicit user decision rather than by passing, and no annotator self-consistency was measured.
Measured on all 3,012 explain rows: closing sentence hedged 3,012/3,012; underscored tokens 0; FINAL lines inside the prose 0; leak hits 0; words 35 / 57 / 77 (min / median / max); sentences 3: 1,759, 4: 1,253; opener top share 5.0 % ("the white screw cap").
Verifier (annotate/D22/verify_goods_explain.py, spec-reimplemented): every row re-derived against the source shard — FINAL == annot byte-exact, both views from the same prose, six columns byte-identical, metadata keys carried, the display-table stamp on every row, judge verdict consistent with the route: 0 problems over 3,115 rows; its planted-fault control (a FINAL moved, an annot changed, a unified-prose byte changed) fails for each planted reason.
Price, measured from the batches' own usage (official Batch list prices of 2026-09-18): smoke $0.17, teacher $6.99, judge $7.44; total $14.60 for 3,115 records. Usage log: outputs/d22_goods_explain/usage_log.jsonl.

Evidence pages

Smoke (24 records, both judge passes, measured prices): reports/D22/goods_explain_smoke.html. Full-run QA (distributions, the pre-filter evolution with every first-pass flag, the judge classes with contact sheets and the reads, the verifier and its control, the billing rows): reports/D22/goods_explain_qa.html. Both are workspace pages, not published.

Loader notes (inherited from D22-goods)

drink_can contains paper cartons as well as cans; food_bottle contains cups, canisters and jars as well as bottles. The category token is the source's label, not a description; the prose names the container by what is seen. 823 of D22's images are MPO (multi-picture JPEG); Pillow decodes the first frame.

Provenance

Two revisions are recorded on purpose. Data: query, image, annot, mask, cate, task, metadata were read from AI4Manufacturing/D22-goods at b4e2f3281653 (2026-09-17; sha-verified shard download) and are carried byte-identically — asserted by the verifier. Card: D22-goods' current revision at the moment of this publish was be217ca810fe (a card-only successor; its 18 parquet files are identical to b4e2f3281653 by LFS sha256 and size, verified at the Hub). Parent canon: AI4Manufacturing/D22 @ 06116762a9b8. Upstream source: PKU-GoodsAD (Zhang et al., IEEE RA-L 2024, arXiv:2307.04956), GPL-3.0; this card is license: other; the repository is gated manual and released for research use.

Generators: annotate/D22/annotate_goods.py (route / batch / pre-filter / judge / assembly), verify_goods_explain.py, qa_evidence.py, publish_goods_annotated.py (forge_model commit recorded by the producer guard at publish).

Downloads last month: 26

Total file size:

6.76 GB

Paper for AI4Manufacturing/D22-goods-annotated

PKU-GoodsAD: A Supermarket Goods Dataset for Unsupervised Anomaly Detection and Segmentation

Paper • 2307.04956 • Published Jul 26, 2023