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

191-region

3,823 region questions: one per gold instance.

The task

One question per gold instance: the query hands the model one box in native pixels (origin top-left) and the closed set of defect types, and the answer is the type in that box.

The query lists the types as display phrases ("pitted surface"); annot is the source token (pitted_surface), which is the machine gold on a rung exactly as on the canon. A scorer maps the model's phrase to the token through common/display_names.

Roles: this is an answer-only tier — annot is the machine-parseable gold AND the training target, in the format the query specifies; reasoning is null by design, because no teacher was called for this repository.

Split, in words

This repository is a MIXED container: it holds records from the canon's train split and from its validation split, so the HF split is named train and the truth for every record is its own metadata.split and metadata.eval_lock.

where the record comes from	records
canon `train`	3,765
canon `validation`	58
total	3,823

load_dataset(...)["train"] therefore hands back train 3,765 + validation 58. A record whose metadata.eval_lock.locked is true is evaluation material wherever it is read; common/overlap.py classifies every record in this repository with 0 unknown, so the container is checkable by the reader rather than promised here.

Overlap / de-duplication (§8)

Cross-family evaluation lock — metadata.eval_lock (manifest revision fe6e286912b0, generated 2026-09-08). Every record of this repository, locked or not, carries metadata.eval_lock, computed by forge_model/common/overlap.py::Overlap.stamp_for against common/overlap_manifest.json at that revision — so within this repository the absence of the key cannot occur. Shape: {"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}. locked is true when the image is evaluation material anywhere in the corpus; own_split marks a record locked by its own split. The per-record field is the authority — the count here is quoted once, at this revision, and a later manifest may change it: 58 of 3,823 records (28 distinct images) are locked — by column: 0 by the cross-family manifest, 58 by their own split. In words: this repository is a MIXED container published under the HF split name train; the 58 locked records are the canon's validation records, and a training pool built from AI4Manufacturing/191-region must exclude them — Overlap.partition / assert_train_pool_clean read the field and will. A stamp whose manifest_revision differs from the current manifest is stale, not wrong — recompute it (Overlap.stamp_is_current); a record with no stamp has not been checked against the corpus as it now is.

⚠ A box-only shortcut exists, and it is disclosed here

A model that never looks at the pixels can guess the type from the box alone. Measured at Stage 4 on this source's geometry, grouped by record so no sibling box leaks:

probe	score
strongest pixel-blind decoder (`gbm_size+position[x,y,w,h,W,H]`)	0.6962
majority baseline	0.2413
kill bar (max(majority + 25 pts, 75%))	0.75

The rung survives the rule — 0.6962 is below the 0.75 bar — but by 5.4 points, so the shortcut is real and is stated rather than discovered later. Type recall from the box alone, per class:

class	recall from geometry alone
`inclusion`	0.829
`scratches`	0.765
`crazing`	0.749
`pitted_surface`	0.741
`patches`	0.659
`rolled-in_scale`	0.387

Consequence for a user of this rung: a score on it is not evidence that a model can SEE the defect type unless it is compared against a geometry-only baseline. The size and position of the box carry most of inclusion, scratches, crazing and pitted_surface; rolled-in_scale is the one class the geometry barely predicts. Source: reports/191/_build/inputs/stage4_screens.json.

Legibility floor

The corpus floor is 16 px rendered min side, measured at the resolution the model sees: 200x200 native renders to 196x196 (scale 0.98), so the 2.36 MP cap never binds here and a 16.3 px native box is the boundary. The unit is per-INSTANCE, not per-record-largest.

Every shipped instance clears the floor: 357 of the canon's instances are below it and are not asked about.

This illegibility is intrinsic to the source: 200 px images cannot be re-rendered larger, so no resolution policy changes it. Sub-floor shares by class are on the canon's own card.

How this repository was checked

A verifier battery reads the canon and the committed query pools and knows nothing about how the builder works, so it and the builder can only agree by both agreeing with the canon. On this rung:

check	result
records	3,823
problems	0
template coverage	191/region 37/37
planted faults caught	10/10

The gold of every record was re-derived from the canon and required to equal what shipped; every query was required to be its pool's template at its recorded index; every record's eval_lock was recomputed; and no query lets a class phrase govern an inflecting verb (common/agreement_probe). Evidence: reports/191/stage5_rungs.html, reports/191/stage5_catalog.html.

Provenance

Derived from AI4Manufacturing/191 at revision eff72aff931d — the canon's own images and human-drawn boxes, unchanged. No model produced any part of this repository: every answer is a function of the canon's geometry.

what	value
canon revision	`eff72aff931d675718c02635110e8b0d8fb3f1d2`
canon shard sha256 (train)	`5db03a541430a3d4f1fd2342c8461e9a0e2f990085ac5bff31f4fcd52af347f4`
canon shard sha256 (validation)	`e8c2aed498a59e71b48457528044aa31ad8dca1a365a3c3820403b9dc154e62b`
builder	`forge_model/annotate/191/build_region.py` (`7a9a34e3bc23275f`)
salt	`191-rungs-v1/region`
cost	$0 · 0 hosted-model calls

query pool	templates	`pool_sha16`
`191/region`	37	`ef9147d431f5b1c0`

Source: NEU-DET (Song & Yan, Applied Surface Science 2013; the NEU surface defect database), packaged by the third-party GitHub project siddhartamukherjee/NEU-DET-Steel-Surface-Defect-Detection (id 361085234, license null). The source tree ships no LICENSE file and the README calls the images NEU's "open surface defect database", so license: other (research use) is the strongest claim this data supports.

Loading

from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/191-region")["train"]          # a MIXED container — see "Split, in words"
import json
md = json.loads(ds[0]["metadata"])
md["split"], md["eval_lock"]["locked"]              # the truth for THIS record

Access is gated manual.

Downloads last month: -

Total file size:

26 MB