datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
elements_annotated_tables_4500_docs
Dataset
🚀 Progress
Last update (UTC): 2025-11-11 15:40:21Z
Documents processed: 4500 / 500058
Batches completed: 30
Total pages/rows uploaded: 89882
Latest batch summary
Batch index: 30
Docs in batch: 150
Pages/rows added: 1487
COCO-Wholebody-annotatedai2thor-perspective-qa-annotated-411-splitsMSP_POD_Annotated_V4Deepface_Annotated_3K
Deepface Annotated 3K Dataset
Deepface_Annotated_3K is a synthetic facial image dataset containing 3K AI-generated faces from StyleGAN2.Each image is automatically annotated with demographic attributes like:
Age (in years)
Gender (with prediction confidence)
Dominant Race (White, Asian, Latino Hispanic, Indian, etc.)
The dataset is designed for research on fairness, bias detection, demographic classification, and synthetic face representation.
Structure information… See the full description on the dataset page: https://huggingface.co/datasets/Subh775/Deepface_Annotated_3K.D15-annotated
D15-annotated
CoT-annotated multi-label defect detection & typing — 2,684 records (train=2684) of the corrected
AI4Manufacturing/D15 (DefectSpectrum), with
teacher-written reasoning (reasoning).
The query/answer formats were redesigned for foundation-model training (see below); the original D15
query/annot strings are not reused.
The repository name is an internal task code. See Provenance below.
Query diversity (2026-07-11). The query field is drawn from a pool of 40 surface… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D15-annotated.193-annotated
193-annotated
Chain-of-thought (CoT) reasoning annotations for Severstal steel-sheet surface inspection —
a binary good / anomalous surface-QC task over 12,568 items (all train), the
reasoning-augmented sibling of AI4Manufacturing/193.
Every image, annot, and mask is byte-identical to the base repo; this repo adds the teacher
reasoning trace, a setting-conditioned query, and a per-record metadata.cot provenance block.
Task
Grade each cold-rolled steel-strip… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/193-annotated.181-annotated
181-annotated
CoT-annotated binary anomaly detection with coarse localization — 6,300 records from
AI4Manufacturing/181 (DAGM2007): all 2,100
anomalous (both splits) + 4,200 goods (2× per class × split, deterministic id-ordered sample of
the 14,000 — a documented cap, chosen because thousands of near-identical "looks uniform" CoTs teach
template memorization, not inspection).
Query diversity (2026-07-11). The query field is drawn from a pool of 40 surface variants for this task… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/181-annotated.D23-annotated
D23-annotated
CoT-annotated defect detection & classification (VISION), two reasoning channels. Derived from
AI4Manufacturing/D23; category B, task T-B2.
The repository name is an internal task code. See Provenance.
Records
1,217 records (train=547 · validation=670) over 8 of VISION's 14 subsets: Cable (172),
Casting (105), Cylinder (283), Electronics (67), Hemisphere (219), Lens (129), PCB_1 (87),
PCB_2 (155). Splits mirror the official VISION train/val; both… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D23-annotated.179-annotated
179-annotated
Chain-of-thought (CoT) reasoning annotations for aero-engine turbine-blade defect inspection (AeBAD)
— 2,160 items (1,011 good + 1,149 defective), the reasoning-augmented sibling of
179-grounding /
179-region /
179-mcq, derived from
AI4Manufacturing/179.
Task
Grade each blade good or defective; if defective, name every defect type and its coarse region.
Defect classes: ablation, breakdown, fracture, groove.
Composition
2,160 rows —… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/179-annotated.D20-goods-annotated-linked-to-D15
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… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D20-goods-annotated-linked-to-D15.186-annotated
186-annotated
Chain-of-thought (CoT) reasoning annotations for magnetic-tile surface-defect inspection — 1,340
items (952 good + 388 defective), the reasoning-augmented sibling of
186-grounding /
186-region /
186-mcq, derived from
AI4Manufacturing/186.
Task
Grade each tile good or defective; if defective, name every defect type and its coarse region.
Defect classes: Blowhole, Break, Crack, Fray, Uneven.
Composition
1,340 rows — good 952 · defective… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/186-annotated.Radiology_Project_Annotatedrefchartqa-balanced-10k-gt-annotatedD11-QA-annotated
Roles
Roles: reasoning (teacher prose) and reasoning_grounded (deterministic, box-cited, generated in code from the gold geometry) are two reasoning registers; either is a valid SFT imitation target and the two count as one view of the record in a mixture. annot.answer is the machine-parseable gold used for verification and reward parsing; the FINAL ANSWER format is the one the query requests, and annot is not an output-format target.
D11-QA — warehouse scene VQA… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D11-QA-annotated.uncut-cells-all-annotatedD05-annotated
D05 Annotated — the full MMAD chain-of-thought corpus
The complete reasoning-augmented version of AI4Manufacturing/D05
(MMAD, multimodal industrial anomaly-detection MCQ): 34,414 records, every one with a
teacher-written chain-of-thought ending in FINAL ANSWER. This is the merge of two tracks
that were built (and are also released) separately, because they make different claims about the
answer:
Track
Records
Answer (annot)
Also released as
Track 1 — rationalized
31… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D05-annotated.invoice-annotated-bboxManually annotated invoice page images exported from AnnotateEverything, with axis-aligned bounding boxes for 8 document-layout regions. Built for training object detectors (YOLO, DETR, etc.) on invoice macro-structure.
Dataset summary
Property
Value
Pages
76
Documents
1
Source PDF
train_images.pdf
Total annotations
771
Avg boxes / page
10.14
Image width range
425 – 2853 px
Image height range
570 – 4096 px
Export date
2026-06-22T19:27:38.375Z… See the full description on the dataset page: https://huggingface.co/datasets/AvoCahDoe/invoice-annotated-bbox.D05-2a-annotated
D05 Track-2a Annotated — re-solved defect classification with chain-of-thought
Want the full corpus in one download? This track is also released merged with
D05-1-annotated as
AI4Manufacturing/D05-annotated
(34,414 records; filter tracks via metadata.cot.source).
A reasoning-augmented, label-corrected subset of AI4Manufacturing/D05
(MMAD, multimodal industrial anomaly-detection MCQ). This release covers Track 2a:
the Defect-Classification questions on the label-poor MMAD… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D05-2a-annotated.D14-codegen-annotated
D14 CodeGen Annotated — CAD code generation with chain-of-thought
The code_gen layer of D14 (BenchCAD) with teacher-written reasoning: 17,900 records, each
4-view orthographic render → reasoning → complete **CadQuery** program. The reasoning ends with the complete program in a ```python block, and the program
is AST-equivalent to BenchCAD's execution-verified reference — asserted per record at assembly
(records whose regeneration drifted carry the reference spliced verbatim;… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D14-codegen-annotated.refchartqa-balanced-1k-gt-annotated-minirps_annotatedD22-goods-annotated
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… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/D22-goods-annotated.ED-D08-annotated
ED-D08-annotated
Chain-of-thought annotated version of AI4Manufacturing/ED-D08 (EngDesign): engineering-design tasks with the reasoning field filled by a teacher LLM. Unified schema, category E, task T-E2. English.
⚠️ Confidence-filtered SUBSET (87 of 101)
EngDesign is a design benchmark whose deliverable varies per task (defined by metadata.output_structure_py) and many tasks admit multiple valid solutions. This release keeps only what we are confident in:… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/ED-D08-annotated.CWRU-annotated
Roles
Roles: reasoning view of CWRU — annot is the source label (ball / inner_race / normal / outer_race), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the image is an envelope spectrum with the theoretical fault frequencies marked. The reasoning column is filled on all 690 records and is the SFT imitation target for this repo; the query enumerates the closed set of labels the answer must come from, and annot… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/CWRU-annotated.annotated_hands_good_dataset
Dataset Card for "annotated_hands_good_dataset"
More Information needed
XJTU-annotated
Roles
Roles: reasoning view of XJTU — annot is the source label (inner_race / normal / outer_race), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the image is an envelope spectrum with the theoretical fault frequencies marked. The reasoning column is filled on all 907 records and is the SFT imitation target for this repo; the query enumerates the closed set of labels the answer must come from, and annot remains the… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/XJTU-annotated.PADERBORN-annotated
Roles
Roles: reasoning view of PADERBORN — annot is the source label (healthy / inner_race / outer_race), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the image is an envelope spectrum with the theoretical fault frequencies marked. The reasoning column is filled on all 710 records and is the SFT imitation target for this repo; the query enumerates the closed set of labels the answer must come from, and annot remains… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/PADERBORN-annotated.IMS-annotated
Roles
Roles: reasoning view of IMS — annot is the source label (ball / inner_race / normal / outer_race), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, where the image is an envelope spectrum with the theoretical fault frequencies marked. The reasoning column is filled on all 542 records and is the SFT imitation target for this repo; the query enumerates the closed set of labels the answer must come from, and annot remains… See the full description on the dataset page: https://huggingface.co/datasets/AI4Manufacturing/IMS-annotated.nayana_bench_human_annotated