| --- |
| tags: |
| - smart-manufacturing |
| - sft |
| - industrial |
| - vision |
| license: other |
| pretty_name: 181-region |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| extra_gated_prompt: >- |
| This dataset is released for **research use**. Access is reviewed and granted |
| **manually** by the maintainers. Please state your name, affiliation, and intended use. |
| --- |
| |
| # 181-region |
|
|
| Region-conditioned binary anomaly QA — **7,548 items**, deterministic (no LLM): 2,100 positives (the |
| ellipse region of every anomalous record; overlay ring or native-px bbox-in-query, ~50/50) + 5,448 |
| clean regions ("no anomaly"; on-image cleans for anomalous records — DAGM has exactly one anomaly per |
| defective image, so off-ellipse regions are genuinely clean — plus sampled good-image regions). Clean |
| boxes sample size AND position from the emitted-positive population; large boxes (≥10% of image) |
| place cleans at MIRRORED pool positions so border-flushness is reproduced — an adversarial review |
| found Class6's border-flush defects vs interior cleans were separable at AUC 0.94 before this fix; |
| after it, per-(class,mode) GBM gates (0.75) show worst cell 0.625 and pooled AUC 0.52/0.53. |
| Machine-verified: zero structural out-of-range rules; every template occurs with both golds; weak |
| residual single-feature signals in the Class6 large-box stratum (~0.58, mechanisms understood) are |
| disclosed. **Clean-majority ratio (72%) is disclosed** — reweight at training time if you want balance. |
|
|
| **Weak-GT disclosure.** DAGM's official labels are deliberately COARSE ellipses ("roughly indicating" |
| the defect) — every localization here is a **containing region**, not a tight extent |
| (`metadata.coarse_gt: true`). Grade localization by containment/center-hit, never tight IoU. |
| > **Query diversity (2026-07-11).** The `query` field is drawn from a pool of **40 per mode (80) surface variants** for this task (paraphrases that preserve the task and answer-format exactly; the answer-format directive is held verbatim), each selected by an independent per-record hash. This replaces the earlier 4-template design to prevent instruction-format overfitting; answers, images, ids, and all provenance are unchanged. A machine gate in `verify_*.py` checks that no template correlates with the gold (binomial z < 4.5). |
| |
| ## Roles |
| |
| **Roles:** this is an answer-only tier — there is no reasoning column; `annot` is both the machine-parseable gold AND the direct-answer SFT target ('SFT-ready' here means direct imitation of `annot` in the query-specified format); it is also the exact-match/IoU reward key for RLVR. |
| |
| ## Overlap / de-duplication (§8) |
| |
| 270 of these images (all anomalous, DAGM Classes covered by DefectSpectrum) also appear byte-identical |
| in the [`D15`](https://huggingface.co/datasets/AI4Manufacturing/D15) family |
| ([`D15-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D15-annotated) / |
| [`D15-mcq`](https://huggingface.co/datasets/AI4Manufacturing/D15-mcq) / |
| [`D15-region`](https://huggingface.co/datasets/AI4Manufacturing/D15-region) / |
| [`D15-grounding`](https://huggingface.co/datasets/AI4Manufacturing/D15-grounding)) with FINE masks and |
| defect-type labels. Reconstruct the exact overlap via `metadata.image_sha256`. **Both official DAGM |
| splits are processed identically here** (project policy): `metadata.split` preserves the original |
| Train/Test membership — carve your own held-out set downstream and keep it out of training. |
|
|
| ## Provenance |
| Built from [`AI4Manufacturing/181`](https://huggingface.co/datasets/AI4Manufacturing/181) by |
| `annotate/181/build_181_derived.py` (forge_model), verified by `verify_181.py`. Exact-match golds |
| (`anomaly` / `no anomaly`). Companions: [`181-annotated`](https://huggingface.co/datasets/AI4Manufacturing/181-annotated), |
| [`181-mcq`](https://huggingface.co/datasets/AI4Manufacturing/181-mcq), |
| [`181-grounding`](https://huggingface.co/datasets/AI4Manufacturing/181-grounding). |
|
|
