| --- |
| tags: |
| - smart-manufacturing |
| - sft |
| - industrial |
| - vision |
| license: other |
| pretty_name: D26 |
| 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. |
| --- |
| |
| # D26 |
|
|
| Binary anomaly detection (public test; private GT withheld -> dropped). Category **B**, task **T-B1**, in the unified Smart-Manufacturing SFT schema. |
|
|
| > The repository name is an internal task code. See **Provenance** below for the underlying dataset. |
|
|
| ## Records |
|
|
| **3,914** records (test=1084 · train=2528 · validation=302). Pixel masks are embedded as a `mask` image column. |
|
|
| ## 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: the plain-text image-level label `good` or `anomalous` (public MVTec-AD-2 is binary — no fine-grained defect types). Pixel-level localization is a separate, deferred task whose target is the `mask` column — see **Task, mask & 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, mask & split |
|
|
| **What this is.** MVTec-AD-2 (Heckler-Kram et al., arXiv:2503.21622, 2025) — advanced unsupervised industrial |
| anomaly detection & localization over 8 scenarios, emphasizing hard cases (transparent / overlapping objects, |
| dark-field & back-light illumination, extremely small defects). The public data is **binary** at the image |
| level — **good** vs **anomalous** — with **no fine-grained defect types**. |
|
|
| **Query & answer (this repo's SFT task).** `query` is our own instruction template (the raw dataset ships no |
| natural-language question). It asks only whether the object is **good** or **anomalous**; `annot` is the |
| plain-text answer `good` or `anomalous`. **The query does not ask for a pixel mask.** |
|
|
| **Mask (localization ground truth for a separate, deferred task).** Each anomalous image ships a single binary |
| ground-truth mask (`mask` column; `1` = defect, `0` = background); normal images have `mask`=null. A text-output |
| model cannot emit a pixel mask directly, so this release keeps the masks as ground truth but does **not** frame |
| localization as the query task (deferred; masks remain for pixel-level evaluation / segmentation models). |
|
|
| **Lighting conditions.** Each scene is captured under several lighting conditions — the `bad` filename suffixes |
| `regular` / `overexposed` / `underexposed` / `shift_*` are **lighting variants, not defect types** — so MVTec-AD-2 |
| can test robustness to real-world illumination shifts. |
|
|
| **Split & withheld data.** `train` = normal only; `validation` = normal only; `test` = the **public** test set |
| (`test_public`: good + anomalous, masks on the anomalous ones). The **private** test sets (`test_private`, |
| `test_private_mixed`) keep their ground truth on the official evaluation server and are **not** included here |
| (not fabricated). See **Records** for exact counts. |
|
|
| ## Provenance |
|
|
| Underlying dataset: **MVTec-AD-2**. Upstream license: **CC BY-NC-SA 4.0** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion script: `D26/convert_d26.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. |
|
|
