README.md

---
tags:
- smart-manufacturing
- sft
- industrial
- vision
license: other
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.
pretty_name: D23-region
dataset_info:
  features:
  - name: query
    dtype: string
  - name: image
    dtype: image
  - name: annot
    dtype: string
  - name: reasoning
    dtype: 'null'
  - name: cate
    dtype: string
  - name: task
    dtype: string
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_examples: 432
  - name: validation
    num_examples: 616
---

# D23-region

Region-conditioned defect typing on VISION industrial imagery — **1,048 items**
(993 gold instances + 55 clean decoys), derived **deterministically**
(no LLM/teacher) from the human gold boxes of
[`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated).
Exact-match gradable (closed per-subset menus) → SFT and RLVR-ready.

> The repository name is an internal task code. See **Provenance** below.

> **Query diversity.** `query` is drawn from a pool of **32 surface variants** (paraphrases preserving task and answer format; the options clause and answer-format directive are held verbatim), selected by an independent salted per-item hash — the corpus query-template-diversity standard.

## Task

"An operator points at a region — what defect, if any, is there?" One item per
eligible gold **instance**: the instance is marked on the full-frame image by a
**red rectangular outline** (D15-region ring convention: thickness scales with image
size), and the query asks which menu option the outlined region contains. Gold =
the class token verbatim (`annot` exact match).

**Padded outlines (disclosed deviation from D15-region).** Every outline is padded
per side by an independent seeded margin in **[10%, 40%]**
of the box dimension (clamped to the frame) — the outline never coincides with the
gold box, so box-tightness cannot carry class information; the geometry-decoder
battery below is computed on exactly these padded outlines. D15-region instead pads
to a fixed minimum size. There is no `bbox_text` mode here (overlay only).

**Clean decoys (PCB_1 only).** 55 regions (28
train / 27 val; seeded fixed ratio
1:3 decoy:gold, realized 0.3313) with size AND
position sampled from PCB_1's own per-instance positive population (D15-region v5
sampler semantics), placed so the padded outline + ring reach has zero overlap with
every gold box (IoU 0, no containment). Their gold is `no defect`.

**PCB_1 clean-region decoys** rest on the same audit: stage-3 defensibility (parent audit artifact `reports/D23/03_stage3_defensibility.html`, §4) found PCB_1's golds complete — its defects are *deliberately synthesized* onto the boards — which is the trust basis for sampling defect-free regions. (The parent card still hardens PCB_1 *prose* against completeness claims; decoys assert absence only inside the sampled region, which additionally avoids every gold box.)

**Menus.** Menus are per-subset and every option is gold at least once
(never-correct options are excised corpus-wide): Hemisphere's menu is **ternary**
(`Defect-A` fails the discriminability bar below and is excised as gold AND option);
`no defect` appears **only** in the PCB_1 menu — the only place it can be correct —
and only in this rung (in MCQ it would be never-correct).

## Records

**1,048 items** (train=432 ·
validation=616), split preserved verbatim from the parent.

| subset | gold items (train/val) | menu (every option is gold ≥1×) | majority answer (post-decoy) |
|---|---|---|---|
| Hemisphere | 625 (230/395) | `Defect-B`, `Defect-C`, `Defect-D` | Defect-D 53.0% |
| Lens | 202 (90/112) | `Fiber`, `Flash Particle`, `Hole`, `Surface Damage`, `Tear` | Fiber 59.9% |
| PCB_1 | 166 (84/82) + 55 decoys | `missing_hole`, `mouse_bite`, `open_circuit`, `short`, `spur`, `spurious_copper`, `no defect` | missing_hole 43.0% |

- Hemisphere carries 62.9% of the gold records (post-kill concentration —
  see the kills below); its ternary menu prior is the majority share above.
- Lens is Fiber-trimmed (seeded) to hold Fiber at 59.9%.
- Within PCB_1, `no defect` is 24.9% of items; global majority
  (Defect-D) is 31.6% — all
  answer priors are asserted ≤60% per category and ≤75%
  globally **after** decoy injection, so blind-guess ceilings equal the shares above.
- Template collisions (measured on the shipped artifact): 61 item pairs
  share their source image AND byte-identical query text (38
  same-gold / 23 different-gold, over 49 records) —
  benign by construction: paired items always differ in the drawn outline, so the
  image disambiguates; expected under independent per-item template draws on
  multi-instance images.

| field | type | meaning |
|---|---|---|
| `query` | str | names the product, mentions the red outline, lists the closed menu; "answer with the option exactly as written" |
| `image` | Image | full-frame photo (never cropped) with the padded red outline burned in |
| `annot` | str | gold class token, or `no defect` (exact match) |
| `reasoning` | null | none — deterministic derivation |
| `cate` / `task` | str | `B` / `T-B2` (unified schema; rungs keep the parent token) |
| `metadata` | str (JSON) | subset, split, `image_sha256`, `source_record_id`, `instance_index` (**deduplicated** objects list; −1 = decoy), gold + raw `bbox_xywh` (native px), `padded_box_xywh`, `drawn_box_xywh`, `pad_margins`, menu, gates provenance |

## Eligibility & kills (the decision trail)

Pools were frozen by a pre-build gate battery under an adversarial convergence
review; the numbers below are the battery's own measurements.

**Legibility floor.** An instance is eligible only if min(w,h) ≥ 16px at the
2.36MP-equivalent training resolution. Unlearnable fraction of each shipped subset's
parent instances by reference input (long side / area cap):

| subset | 448 | 768 | 1024 | 1568 | 2.36MP |
|---|---|---|---|---|---|
| Hemisphere | 88.1% | 71.9% | 54.2% | 22.4% | 17.4% |
| Lens | 93.4% | 87.8% | 83.5% | 68.1% | 68.8% |
| PCB_1 | 100.0% | 97.5% | 85.6% | 69.7% | 58.1% |

**Geometry-decoder kills** (rule: hard decoder iff probe >= majority + 25pts AND >= 75% absolute, computed on padded
outline geometry — no retained subset's outline geometry decodes its answer):

- **PCB_2 killed** — full-geometry GBM 98.1% vs
  30.5% majority (repeated-panel layout).
- **Cable killed (region/MCQ)** — size-only GBM 82.4%
  vs 50.4% majority: box size alone decodes the class.
- **Casting killed** — full-geometry GBM 83.8%
  vs 56.6% majority; the kill is position-driven
  (position channel adds +10.0pts over size-only
  73.8%).
- **Cylinder killed (round-3)** — cross-split geometry twins: 21.7%
  of val instances have a train instance within 10px (raw geometry) and
  51.3% within 30px; train→val padded 1-NN
  80.3% vs 33.6% val-majority —
  repeated rig positions leak labels across the split.

**Retained-subset certifications** (same probes, below the bar):

- **Hemisphere (BCD)** — cross-split padded 1-NN 71.9%
  (+16.7pts over val-majority) with twin fractions
  0.0%/8.1% —
  the signal is class-geometry, not rig repetition; in-pool padded LOO 1-NN on the
  **shipped** pool = **0.773** vs kill bar
  **0.780** (passes by
  0.7pts; both digits disclosed —
  this is the highest retained value in the corpus).
- **Lens (post-Fiber-trim)** — cross-split padded 1-NN 50.0%
  (-8.0pts), twins
  0.9%/3.6%.
- **PCB_1** — cross-split padded 1-NN 37.8%
  (-26.8pts), twins
  3.7%/7.3%.

**Opaque-code discriminability (size-normalized audit).** Hemisphere's anonymized
`Defect-A..D` codes were additionally required to show appearance-borne signal on a
fixed 224px canvas (removes
absolute size by construction; judge-side instrument, upscaling allowed): Defect-B
3.2× chance, Defect-C 2.0×,
Defect-D 1.73× pass the ≥1.5× bar; **Defect-A
0.8× fails and is excluded as a gold everywhere** (menu
excised too — never-correct options are removed corpus-wide).
The verdicts are **judge-robust** (judge-sensitivity rerun, three control-passing
judges: claude-sonnet-4-5, claude-sonnet-4-6, gpt-5.6): Defect-A stays under the ≥1.5× bar with every judge
(accuracy 0.20/0.37/0.30 vs the 0.375 bar), while Defect-B/C/D and every PCB_2
class pass under all three.
*Instrument-ceiling caveat:* the positive control (Cylinder, semantically named
classes) passes the control gate under every judge (macro 0.65/0.74/0.75), but its
Chip class scores 0.15 under the original judge and rises to
≈0.45 under the stronger judges — per-class collapses are **partly
judge-limited, not a hard pixel ceiling**; treat per-class passes as a lower bound
on discriminability. Residual cue disclosed by the protocol: **aspect ratio
survives the canvas normalization**.

*No D23-counting rung exists:* counting was dropped at plan review (degenerate ~all-one count priors on this parent + the counting shape lacks PoC validation; the shape is kept corpus-wide via other artifacts).

## Roles

**Roles:** this is an answer-only tier — there is no reasoning content (`reasoning` is null on every record); `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.

## Provenance

Derived read-only from [`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated) (revision pinned: `fd728cc34406cd87512bd22fd8c31df78dace149`), itself derived from raw VISION (Bai et al., arXiv:2306.07890; upstream CC BY-NC 4.0 — respect upstream terms; this card is `license: other`). Answers are **pure functions of the human gold annotations** — no LLM/teacher anywhere in this rung. Generator: `annotate/D23/rungs/build_rungs.py` in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model) (builder sha256 `8652395b…`), built against the pre-build gate battery (gates v5 report, script sha256 `0b9d0e43…`, build source of truth = its section 16). Every independent choice is salted-hash seeded; a rerun reproduces the artifact exactly (tuple-hash verified).

Parent golds were deduplicated first: 5 records carried 8 exact-duplicate (class,bbox) phantom instances, collapsed before any pool math. `metadata.instance_index` refers to this **deduplicated** objects list, not the raw parent array. The parent's official train/validation split is preserved verbatim on every item (uniform-split policy; carve train/eval downstream).

## Training-mixture notes

- **One-lineage rule:** the parent excludes the 641 VISION images byte-shared with the DefectSpectrum/D15 family (materially different label policies); the 8 subsets here have **zero** image overlap with the D15 family. Details + machine-readable keys: base [`AI4Manufacturing/D23`](https://huggingface.co/datasets/AI4Manufacturing/D23) card §8.
- **Same-evidence rungs:** 137 of 140 D23-mcq marked instances also ship as D23-region records with **identical padded outlines** — region and mcq are format variants over the same evidence. Treat them jointly in any mixture/eval carve; never split the two rungs across train/eval.
- **Image-wise carving:** one photo can appear in the parent and in up to three rungs; carve on `metadata.image_sha256` across the whole D23 family simultaneously.
- **All-defective world prior:** every parent record is defective; rung items never assert a defect-free image (clean **regions** exist only as PCB_1 decoys here).

## Overlap / de-duplication (§8)

Base photos are the SAME images as [`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated) and base [`AI4Manufacturing/D23`](https://huggingface.co/datasets/AI4Manufacturing/D23) — this rung inherits their overlap situation: the shipped 8-subset lineage shares no imagery with the D15 family (sha-verified sidecar on the base card), and D23-validation imagery appears in the other D23 rungs and the parent. **Do not evaluate on any D23-family repo's validation split if you train on this set**, and reconstruct exact overlaps via `metadata.image_sha256`.

## Training notes

- Loss on completions; answers are short by instruction. Exact-match reward for RLVR.
- 137/140 of
  [`AI4Manufacturing/D23-mcq`](https://huggingface.co/datasets/AI4Manufacturing/D23-mcq)'s marked instances
  are also region records here with identical outlines — same-evidence format
  variants (see Training-mixture notes).
- Companions: [`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated) (CoT channels),
  [`AI4Manufacturing/D23-mcq`](https://huggingface.co/datasets/AI4Manufacturing/D23-mcq),
  [`AI4Manufacturing/D23-grounding`](https://huggingface.co/datasets/AI4Manufacturing/D23-grounding).