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-grounding
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: 121
  - name: validation
    num_examples: 162
---

# D23-grounding

Single-class detection-format defect localization on VISION industrial imagery —
**283 items** (263 positive (record, class) pairs + 20 Cable
negatives), derived **deterministically** (no LLM/teacher) from the human gold
boxes of [`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated). The model outputs
**boxes as text** (absolute-coordinate JSON, D15-grounding format parity).

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

> **Smallness is deliberate.** 283 items is what survives the anti-duplication
> and floor rules below. The rung ships as its own repo because the corpus PoC
> leave-one-out found grounding rungs **load-bearing** (removing them: grounding
> F1 -8.2, in-domain exact -2.8), and per-shape format parity
> with the D15/181 grounding rungs is what makes those findings transfer.

> **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

"Locate every {class}." Each query names ONE class; `annot` is a JSON list of
`{"type": ..., "bbox_xywh": [x, y, w, h]}` in **native pixel coordinates** (origin
top-left, sorted by x then y), one entry per gold instance of that class — or `[]`
when the class is absent (Cable negatives). **No overlay** — the image is the raw
full frame.

- **Anti-duplication rule:** a positive (record, class) pair is eligible only if the
  record contains ≥2 distinct classes — on single-class records the answer would
  equal the parent detection task's full answer verbatim. Consequence: the queried
  class is **always present except on the 20 Cable negatives** — a disclosed
  limitation: within Hemisphere/Lens this rung teaches per-class localization, not
  per-class absence.
- **Pair-level floor:** every instance of the queried class must pass the legibility
  floor (an answer may not silently omit an illegible box).
- **Cable negatives** (20: `break` 11 / `thunderbolt` 9;
  1:1 with Cable positives per split, salted-hash selected): queries about the class
  absent from a single-class Cable record, gold `[]`.
- **Cable absence golds** rest on the exhaustiveness audit, not on assertion: the stage-3 defensibility review (parent audit artifact `reports/D23/03_stage3_defensibility.html`, §4 exhaustiveness spot-check) found Cable's gold sets complete-looking with no unboxed defects, and the parent card's exhaustiveness-tier note places Cable in the completeness-claiming tier.
- *Negatives disclosure:* answering the Cable negatives requires the
  break/thunderbolt distinction whose padded box geometry killed Cable's region
  items; here it is perception-mediated (no overlay) — a split-respecting geometry
  decoder reaches 62.7% on the padded boxes, under the
  75.4% kill bar.

## Records

**283 items** (train=121 · validation=162),
split preserved verbatim from the parent. Global empty-answer share
7.07%.

| subset | items | queried classes (each ≥1×) | max query-class share | empty-answer share |
|---|---|---|---|---|
| Cable | 40 | `break`, `thunderbolt` | 52.5% | 50.0% |
| Hemisphere | 194 | `Defect-B`, `Defect-C`, `Defect-D` | 51.5% | 0.0% |
| Lens | 49 | `Fiber`, `Flash Particle`, `Hole`, `Surface Damage`, `Tear` | 38.8% | 0.0% |

**Exits:** Casting (4 multi-class pairs) and Cylinder
(2) fall below de-minimis once single-class records are
excluded; PCB_1 has 0 multi-class pairs; Electronics is
single-class (anti-duplication); PCB_2 is geometry-killed (see the parent-family
kill table on the region card); Hemisphere `Defect-A` is excluded as a gold
(discriminability audit).

| field | type | meaning |
|---|---|---|
| `query` | str | names the product and ONE class; JSON output spec + empty-list rule held verbatim |
| `image` | Image | raw full-frame photo (no overlays, never cropped) |
| `annot` | str | JSON box list (see above), `[]` on negatives |
| `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`, `query_class`, `n_instances`, `negative`, `gold_boxes_xywh` (native px), 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 |
|---|---|---|---|---|---|
| Cable | 31.2% | 7.3% | 4.0% | 1.6% | 1.2% |
| Hemisphere | 88.1% | 71.9% | 54.2% | 22.4% | 17.4% |
| Lens | 93.4% | 87.8% | 83.5% | 68.1% | 68.8% |

**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 (empty answers here assert only that a *named class* is absent).

## 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

- Boxes are canonical **native-px COCO xywh**. Convert to your model's grounding
  convention at train time (normalized, corner pairs, special tokens, …) —
  regenerate, don't regex; see `common/box_convert.py` in forge_model.
- Gradable by class-aware box matching (e.g. greedy IoU≥0.5) → RLVR reward or eval
  metric; `[]` records score rejection of the named class.
- Companions: [`AI4Manufacturing/D23-annotated`](https://huggingface.co/datasets/AI4Manufacturing/D23-annotated) (CoT channels
  incl. the coordinate-citing `reasoning_grounded`),
  [`AI4Manufacturing/D23-region`](https://huggingface.co/datasets/AI4Manufacturing/D23-region),
  [`AI4Manufacturing/D23-mcq`](https://huggingface.co/datasets/AI4Manufacturing/D23-mcq).