You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
D20-goods-annotated (linked to D15)
Chain-of-thought inspection annotations for the good (defect-free) examples of MVTec-AD (D20) —
2,066 items, all good — assembled as a positive-sample supplement for
AI4Manufacturing/D15-annotated.
Relationship to D15 (why this set exists)
D15-annotated (the DS-MVTec part) is anomaly-heavy — 386 good vs 1,226 defective (~0.31 : 1; the
pill category has zero goods). This set is D15's positive (good) examples: 2,066 MVTec
train-split good images, byte-disjoint from every D15 image (sha256-verified), chosen so that when
concatenated with D15-annotated each category reaches good : anomaly = 2 : 1.
Intended use: mix with D15-annotated. The two share an identical 8-column schema (query drawn
verbatim from D15-annotated's own per-category query pools), so D15-annotated (anomalies) + this set
(goods) concatenate directly into a class-balanced good/anomaly CoT-SFT corpus.
Task
Grade each part good. The CoT scans the surface and rules out defects without inventing any —
every FINAL ANSWER is good. This is exactly the failure mode CoT annotation must get right: not
hallucinating a defect on a clean part.
Composition
2,066 rows, all good, across 15 MVTec categories: bottle 106, cable 126, capsule 214, carpet 150, grid 93, hazelnut 136, leather 152, metal_nut 118, pill 264, screw 197, tile 135, toothbrush 48, transistor 20, wood 101, zipper 206.
Schema (8 columns)
| field | meaning |
|---|---|
query |
student question (prose; states the closed defect-class list) |
image |
the raw inspection photo |
annot |
machine-parseable gold JSON: `{"defects": [{"region", "type"}, ...], "label": "good" |
reasoning |
teacher chain-of-thought ending FINAL ANSWER: ... |
mask |
pixel GT mask (defective rows) |
cate / task |
B / T-B2 |
metadata |
source, category, image_sha256, image_path, split, … |
How the reasoning was made
- Teacher:
gpt-5.4-mini(OpenAI Batch API), English, gold-conditioned rationalization — the teacher writes forward reasoning that lands on the given human gold; it never re-solves the image and never rejection-samples the answer. The gold is authoritative; the reasoning explains it. - No leakage: the prompt forbids naming any grounding artifact (mask / ground-truth / reference / coordinates / "the given answer"); a regex gate confirms 0 leaked-vocabulary hits.
Roles
Roles: reasoning is the SFT imitation target — its FINAL ANSWER segment is the model-facing answer format; annot is the machine-parseable gold used for verification and reward parsing, not an output-format target.
Grounded traces
The reasoning_grounded column was removed 2026-07-21 per the corpus grounded-trace doctrine
(decided 2026-07-18): for classification-shape tasks the trace can only restate the verdict (with
unrequested coordinates) — geometry stays in metadata/mask; traces are pure functions of the GT and the
committed generator (forge_model annotate/D20/) and can be re-rendered on demand. Default SFT
view: query -> reasoning. RLVR uses query + annot.
Faithfulness gate (Stage 7)
- 0 leaked-vocabulary hits across the set.
- 0 hallucinated defects on good images — the key failure mode (a clean part narrated as defective).
- Gold-conditioned faithfulness judge (
gpt-5.6-terra, a different model family from the teacher) over a 304-row good-heavy calibration sample: 1.7% flag rate, 0% on good rows. Every flagged case was either a judge conflation on a correct row or was fixed and re-verified (see the source-specific notes).
Provenance / reproduction
Built from the source dataset's human annotations by
forge_model annotate/ (builders + shared
annotate/teacher_prompt.py + annotate/cot_batch/ pipeline). Public, manual access review
(gated=manual).
Upstream license (inherited)
The images in this dataset are MVTec-AD goods, obtained via
AI4Manufacturing/D20.
Upstream license: CC BY-NC-SA 4.0 (this card is license: other; respect the upstream terms).
This license is inherited, not chosen, and it is viral: both the NonCommercial restriction and the ShareAlike obligation propagate to this dataset and to anything derived from it. Any redistribution or derivative must remain non-commercial and carry the same license plus attribution to MVTec Software GmbH. Mixtures that include these records inherit the same constraint.
- Downloads last month
- 42