| --- |
| tags: |
| - smart-manufacturing |
| - sft |
| - industrial |
| - vision |
| license: other |
| pretty_name: "191" |
| 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. |
| --- |
| |
| # 191 |
|
|
| Steel-surface defect detection (6 classes; VOC bbox). Category **B**, task **T-B2**, in the unified Smart-Manufacturing SFT schema. |
|
|
| > The repository name is an internal task code. See **Provenance** below for the underlying dataset. |
|
|
| ## Records |
|
|
| **1,800** records (train=1770 · validation=30). |
|
|
| ## Unified SFT schema |
|
|
| | field | type | meaning | |
| |---|---|---| |
| | `query` | str | the question / instruction (model input) | |
| | `image` | Image | the input image (bytes embedded) | |
| | `annot` | str | the answer — for this dataset: one `class,[x, y, w, h]` line per defect bounding box (COCO x/y/width/height in pixels; converted from the source Pascal-VOC corner boxes). The 6 steel-defect classes are a closed set given in the query; full boxes + image size are in `metadata.objects`. Detection task — no mask column — see **Task & split** below | |
| | `reasoning` | null | no native CoT in these datasets | |
| | `cate` | "B" | SFT category | |
| | `task` | "T-xx" | unified task id | |
| | `metadata` | str (JSON) | split, provenance, `image_path`, `image_sha256` (dedup key) | |
| | `mask` | Image \| null | *(T-B1/T-B2 only)* the pixel ground-truth mask, bytes embedded | |
| | `masks` | list[Image] | *(D21 only)* multi-region masks | |
|
|
| ## Task & split |
|
|
| **What this is.** NEU-DET — the detection version of the NEU Surface Defect Database (Song & Yan, Northeastern |
| University): 1,800 grayscale 200x200 steel-surface images across **6 defect classes** (crazing, inclusion, |
| patches, pitted_surface, rolled-in_scale, scratches), each annotated with Pascal-VOC bounding boxes (4,126 boxes |
| total). Every image contains at least one defect — there are no defect-free images. |
|
|
| **Task.** Object **detection**: localize and classify every surface defect. `query` (our template) names the |
| closed set of 6 classes and asks for one `class,[x, y, w, h]` line per box (top-left x, y + width, height, in |
| pixels). `annot` is exactly that — the source VOC corner boxes `[xmin,ymin,xmax,ymax]` converted to COCO |
| `[x,y,w,h]`. **There is no mask** — localization is the bounding box. Full boxes + image size are preserved in |
| `metadata.objects` and `metadata.width`/`height`. |
|
|
| **Split.** The source ships a held-out `Validation_Images` folder (30 images, verified 0 overlap with the main |
| set) -> `train` (1,770) + `validation` (30) = 1,800. |
|
|
| ## Provenance |
|
|
| Underlying dataset: **NEU-DET**. Upstream license: **other (research use; NEU Surface Defect Database)** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion script: `191/convert_d91.py`, published with `publish/push_to_hf.py`, both in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model). |
|
|
| ## Overlap / de-duplication (§8) |
|
|
| None notable. Each record carries `metadata.image_sha256` so overlapping images can be kept entirely on one side of a train/eval split. |
|
|
