README.md

---
tags:
- smart-manufacturing
- sft
- industrial
- vision
license: other
pretty_name: "D20-goods"
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.
---

# D20-goods

**2,066 good (defect-free) MVTec-AD images** formatted as a drop-in **class-balance supplement for
[`D15-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D15-annotated)**.

> The repository name is an internal task code: images come from `AI4Manufacturing/D20` (MVTec-AD),
> formatting follows `D15-annotated` (DefectSpectrum re-annotation of the same product families).

## Why

D15-annotated's DS-MVTec portion is anomaly-heavy — 386 good vs 1,226 defective overall (0.31:1),
with `pill` at **zero** goods and `capsule`/`hazelnut` at 4 each. A model trained on that mix gets
almost no signal for the "declare good" decision. Mixing this set in brings **every category to
good:anom = 2:1** (the ratio used by the in-house `181-annotated` precedent):

| category | D15 anom | D15 good | added here | category | D15 anom | D15 good | added here |
|---|---|---|---|---|---|---|---|
| bottle | 63 | 20 | 106 | pill | 132 | 0 | 264 |
| cable | 92 | 58 | 126 | screw | 119 | 41 | 197 |
| capsule | 109 | 4 | 214 | tile | 84 | 33 | 135 |
| carpet | 89 | 28 | 150 | toothbrush | 30 | 12 | 48 |
| grid | 57 | 21 | 93 | transistor | 40 | 60 | 20 |
| hazelnut | 70 | 4 | 136 | wood | 60 | 19 | 101 |
| leather | 92 | 32 | 152 | zipper | 119 | 32 | 206 |
| metal_nut | 70 | 22 | 118 | **total** | **1,226** | **386** | **2,066** |

(D15 anomalous counts above exclude nothing; MVTec's `flip` and `pill_type` classes are absent from
D15 by DefectSpectrum's own design and are NOT covered by this supplement either.)

## Format compatibility (mixes as ONE training set with D15-annotated)

- **`query`** — drawn **verbatim** from D15-annotated's own per-category query-variant pools
  (harvested from the published dataset; 26–40 variants per category), selected by per-record hash.
- **`annot`** — byte-identical to D15-annotated's good gold: `{"defects": [], "label": "good"}`.
- **Schema** — identical columns: `query, image, annot, reasoning, cate, task, mask, metadata`
  (`cate=B`, `task=T-B2`).
- **`reasoning` is EMPTY in this release** — the CoT pass is scheduled separately and will land as
  a dataset revision. `mask` is empty (good rows have no defect mask).
- **Roles** (once the CoT lands): `reasoning` is the SFT imitation target; `annot` is the
  machine-parseable gold / reward key, **not** an output-format target.
- The empty `reasoning_grounded` column was removed 2026-07-21 per the corpus grounded-trace
  doctrine (decided 2026-07-18): classification-shape tasks never carry the column — geometry stays
  in `metadata`/`mask`; no grounded chains will be added. This matches `D15-annotated` and
  `D20-goods-annotated-linked-to-D15`, which had the column removed under the same doctrine.

## Provenance & disjointness

Images are the **MVTec-AD train-split goods** taken from `AI4Manufacturing/D20` (revision `69b0409`);
metadata carries `image_sha256`, the original `image_path`, and a `supplement` block (per-category
targets + sampling doc). **Byte-disjointness from every D15 image is machine-verified by sha256 at
build time** (D15 only ever ingested MVTec *test* images; the 4 known D20↔D15 shared-pixel families
are disclosed in the D20/D15 repos). Sampling is a deterministic even spread over path-sorted train
goods per category, which covers the known appearance variants (pill speckle-density range,
toothbrush bristle colors blue/red/yellow, hazelnut per-nut variation).

Generator: `annotate/D20/build_d20_goods.py` (forge_model); compat assets harvested into
`annotate/D20/d15_compat.json`; verification gates in the same script (`verify`).

## Upstream license (inherited)

The images in this dataset are **MVTec-AD** goods, obtained via
[`AI4Manufacturing/D20`](https://huggingface.co/datasets/AI4Manufacturing/D20).
Upstream license: **CC BY-NC-SA 4.0** (this card is `license: other`; respect the upstream terms).

This license is **inherited, not chosen**, and it is **viral**: both the **NonCommercial** restriction
and the **ShareAlike** obligation propagate to this dataset and to anything derived from it. Any
redistribution or derivative must remain non-commercial and carry the same license plus attribution to
MVTec Software GmbH. Mixtures that include these records inherit the same constraint.