README.md

---
tags:
- smart-manufacturing
- sft
- industrial
- vision
license: other
pretty_name: "218"
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.
---


<!-- ROLES-CANON:BEGIN -->
## Roles

**Roles:** canon repo — `annot` is the source label, kept machine-parseable as the gold for verification and reward parsing; there is no filled `reasoning` column and this repo is **not itself a training view**. Derived repos each state their own regime on their own card.
<!-- ROLES-CANON:END -->
# 218

Image-level aesthetic-defect classification of laser weld seams on EV battery cells (61,670 narrow strips; binary). 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

**61,670** records (test=9251 · train=43169 · val=9250). 

## Unified SFT schema

| field | type | meaning |
|---|---|---|
| `query` | str | the question / instruction (model input) |
| `image` | Image | the input image (bytes embedded); for multi-image rows, a preview of the first view |
| `images` | list[Image] | *(multi-image rows)* all input views / modalities for the row, bytes embedded |
| `annot` | str | the answer — for this dataset: one word, `ok` or `defective`. The deposit's own folder token (`OK` / `KO`) is preserved in `metadata.source_label`; `KO` is not self-explanatory so it is not the answer string. **⚠ Report balanced accuracy or defect recall, never plain accuracy** — a single threshold on the image's WIDTH reaches 81.3% balanced accuracy with no pixels read, and its plain accuracy (75.4%) sits below the majority floor, which hides it. See the card |
| `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] | *(multi-image T-B1 / D21)* per-view masks aligned with `images` (None where a view has no defect), or multi-region masks |

## Licence

**CC BY 4.0.** Read from the Zenodo record's own `metadata.license` field at fetch time, not inferred
from a paper. Redistribution is permitted with attribution and no permission request was needed.

Cite the deposit: *"2D image dataset for the presence or absence of aesthetical defects on EV battery
welds"*, OPENZDM project, Zenodo, DOI
[10.5281/zenodo.14025630](https://doi.org/10.5281/zenodo.14025630). The archive
(`Weld Aesthetical Defects.zip`, 1,153,183,844 B) was verified against Zenodo's **published md5**,
recomputed locally, before anything was built from it.

## ⚠⚠ Read this before reporting any number: the image's WIDTH carries most of the label

`KO` strips are almost all ~400 px wide (p10 388); `ok` strips span 48–651 px (p10 92). So a single
threshold on the image's width — **reading no pixels at all** — recovers most of the defect class:

| pixel-blind predictor | accuracy | defect recall | specificity | **balanced accuracy** |
|---|---:|---:|---:|---:|
| always the majority class (`ok`) | 80.9% | 0.0% | 100.0% | 50.0% |
| **`defective` iff width ≥ 350 — one threshold** | 75.4% | **90.8%** | 71.8% | **81.3%** |
| size-bin lookup (width/10, height) | 83.0% | 36.1% | 94.0% | 65.1% |
| intensity-bin lookup (mean, std) | 81.1% | 3.6% | 99.4% | 51.5% |
| **the image's width and height alone** — the standing corpus floor (class-balanced linear probe, no pixels read; `common/lazy_floors.py`) | 75.4% | 90.8% | 71.8% | **81.3%** |

The last row is the corpus's standing frame-size floor, computed on every image dataset from the
build: fitted on train, scored on test, 3,355 distinct frame sizes. It **rediscovers the one-threshold
rule exactly** (its weight sits almost entirely on the width), which is the check that the two
instruments agree.

**Report results as balanced accuracy or defect recall. Plain accuracy hides this shortcut
completely.** The width rule's plain accuracy is **75.4%, which is *below* the 80.9% majority
floor** — read that column alone and you conclude the shortcut is worthless. It is not: on balanced
accuracy it is 81.3% against a 50% chance line and it recovers 90.8% of every defective strip.

*This is not a hypothetical caution.* The first pass of this dataset's own audit scored the probes on
plain accuracy, saw "+2.1 pp over the floor", and concluded the shortcut was minor. That conclusion
was wrong, and it was wrong because of the metric, not the measurement.

The intensity probe reaches only 51.5% balanced, so the artefact is **geometric, not photometric** —
not exposure or brightness. Defect rate by width:

| width | images | defective |
|---|---:|---:|
| 50–99 | 9,603 | **0.2%** |
| 100–149 | 1,307 | 0.8% |
| 300–349 | 25,660 | 4.2% |
| 350–399 | 4,011 | 36.8% |
| 400–449 | 20,648 | **44.4%** |
| **width < 350** | **36,997** (60.0%) | **3.1%** |
| **width ≥ 350** | **24,673** (40.0%) | **43.2%** |

A 14× shift in the class prior from one number that needs no pixels. The deposit records nothing that
would explain it; the likeliest cause is that short crops come from a different product, station or
seam length whose failure rate is simply lower.

**Nothing is filtered or re-cropped to hide this.** Canon measures; the mixture decides the dose.
Every record carries `metadata.image_wh`, so the artefact is auditable per record rather than only in
aggregate.

## Adapt-layer instruction

Two consequences, stated so the choice is made deliberately rather than by a default:

1. **Any resize to a fixed size — padding included — changes or destroys the width shortcut.** That
   may be what you want, but it is a change to what the task measures, not a formatting detail.
2. **Any view that preserves native width hands the shortcut to the model.**

Strip geometry: width median 323 (p10 94, p90 408), height median **37** (min 22, max 70), median
aspect ratio **9.4:1**. A square-ish resize either pads the strip — spending most of the token budget
on blank space — or squashes it, destroying the along-seam texture the classes differ in. A view
should render at **≥3× height (37 → 111 px) and preserve the aspect ratio**. The choice belongs to
layer 3; this card states the consequences only.

## Records and split

| split | `ok` | `defective` | total | defective share |
|---|---:|---:|---:|---:|
| train | 34,916 | 8,253 | 43,169 | 19.1% |
| val | 7,482 | 1,768 | 9,250 | 19.1% |
| test | 7,482 | 1,769 | 9,251 | 19.1% |
| **total** | **49,880** | **11,790** | **61,670** | 19.1% |

**The authors' own 3-way split is adopted unchanged** and recorded in `metadata.official_split`. It
is stratified to a 19.1% defect share in all three parts, to one decimal.

⚠ The middle split is named **`val`**, not `validation` — that is the deposit's own folder name and it
is kept verbatim, so `load_dataset(...)["val"]` is the key that works.

### Leakage: no evidence, which is not the same as proof

The deposit has **no capture id** and filenames restart at `image_1000.png` inside every split/class,
so nothing states which strips came from the same weld. Image similarity is the only available
evidence, and it was measured against a control:

| comparison | exact match | within Hamming 4 |
|---|---:|---:|
| test → train (the question) | 0.00% | **0.18%** |
| trainA → trainB (an arbitrary boundary inside train, **control**) | 0.00% | 0.02% |
| | | **excess +0.17 pp** |

The control is a boundary that by construction does *not* separate welds, so whatever rate it
produces is the hash's background collision rate on this data. An excess of a sixth of a percentage
point is nothing.

⚠ A first measurement, using a standard square 8×8 difference hash, reported **9.0%** and looked like
serious leakage. That was the instrument: squashing a 400×35 strip into a 9×8 thumbnail throws away
the along-seam detail that distinguishes one weld from another, so unrelated strips collided. A
strip-shaped **16×4** hash gives the 0.18% above.

⚠ Stated precisely: this is **no evidence of leakage**, not **proof of a clean split**. A weld imaged
twice from different angles would not look similar and would not be caught by any image-similarity
test. With no capture id in the deposit, that possibility cannot be excluded by anything.

## Duplicates: clean

**61,670 distinct sha256 over 61,670 images** — zero duplicate copies, zero groups spanning more than
one split, zero label conflicts. Worth stating because it is not the norm: `207` shipped 120 duplicate
hashes with 79 label conflicts, and `096` was unusable partly for that reason.

## What this dataset does not contain

⚠ **No boxes and no masks.** A defect-size distribution **cannot be computed** for this dataset, and
anything quoting one is inventing it. The label is image-level only, so the corpus's 16 px legibility
floor — which gates *localization* — does not bind here: nothing has to be resolved to a box.

There is no detection rung to build from this source.

## Version history

**v1 — published 2026-09-09.** 61,670 records (43,169 train / 9,250 val / 9,251 test), one query
template.

**v2 — this revision. Query text only.** Every record's query is drawn from a pool of **37
gate-verified paraphrases** of the v1 wording, assigned by a stable hash of the image path and recorded
as `metadata.query_template`; **index 0 is v1's wording, byte for byte** (1,677 records keep it).
Answers, images, split and every other field are unchanged. The pool passed the deterministic gates
(names the target, does not misdescribe the strip, keeps the two-word answer form, introduces no
class name the source does not state) and a template ↔ gold independence test on this build (37
templates, 0 flagged). Two more things land with it: the **build-time pixel-identity check** (§8 now
carries measured numbers: 61,670 images, 61,670 distinct, 0 on both sides of any split — the same
result the sha256 count above reached, now on decoded pixels) and the **frame-size floor** row in
the pixel-blind table, which reproduces the width rule.

## Provenance

Underlying dataset: **OPENZDM EV battery welds**. Upstream license: **CC BY 4.0 (Zenodo record metadata.license; OPENZDM project) — redistribution permitted with attribution, no permission request needed** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under `218/` (with `publish/push_to_hf.py`) in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model).

Converter: `forge_model@cae2e84` (PR #98). That is the last commit to touch this dataset's converter, which is what produced the data; this card's own text lives in `publish/push_to_hf.py` and moves independently.

## Overlap / de-duplication (§8)

No overlap with any other dataset in this corpus — laser welding of battery interconnects, a different process from `216` (MAG bead) and `102`/`214` (radiography), and 61,670 distinct sha256 with zero duplicates. ⚠ The deposit carries **no capture id**, so nothing states which strips share a weld; the official 3-way split is adopted and the leakage evidence is on the card.

**Two identities, and they answer different questions.** `metadata.image_sha256` hashes the **file bytes**: it finds byte-identical copies and is blind to a re-encode. `metadata.pixel_sha256` hashes the **decoded image** (`mode | size | pixels`): it finds the same photograph saved twice. Only the second one settles whether an image is duplicated.

Measured **at build time**, not asserted afterwards — a violation aborts the build and names the offending records:

| | |
|---|---:|
| images checked | 61,670 |
| distinct by decoded pixels | 61,670 |
| images carrying more than one record | 0 |
| images on **both sides** of the split | **0** |

<!-- EVAL-LOCK:BEGIN -->
**Cross-family evaluation lock — `metadata.eval_lock` (stamped 2026-09-20; manifest revision `fe6e286912b0`, generated 2026-09-08).** Every record of this repository, locked or not, carries `metadata.eval_lock`, computed by `forge_model/common/overlap.py::Overlap.stamp_for` against `common/overlap_manifest.json` at that revision — so within this repository the absence of the key cannot occur. Shape: `{"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}`. `locked` is true when the image is evaluation material anywhere in the corpus; `against` names every repository and split in which it is (sorted; `[]` when not locked; it includes the record's own family where that is so); `own_split` marks a record locked by its own split. **The per-record field is the authority** — the count here is quoted once, at this revision, and a later manifest may change it: **18,501 of 61,670 records (18,501 distinct images) are locked** — by column: **0 by the cross-family manifest, 18,501 by their own `split`**, 0 both ways and counted once; counterparts (records per counterpart; a record can appear under several): none — every lock here is by the record's own split; 9,251 locked by their own `split: test`, 9,250 locked by their own `split: val`. **In words: 18,501 of the 61,670 records in this repository are evaluation material by their own `metadata.split` (`test`: 9,251, `val`: 9,250) and sit inside the HF split named `test` / `train` / `val` — under the uniform-split convention the HF split name is a container name, and `metadata.split` together with `metadata.eval_lock` carries the truth; a train pool must exclude them**. A stamp whose `manifest_revision` differs from the current manifest is *stale*, not wrong — recompute it (`Overlap.stamp_is_current`); a record with no stamp has not been checked against the corpus as it now is. `Overlap.partition` / `assert_train_pool_clean` read the field: a train pool built from this repository must exclude every locked record.
<!-- EVAL-LOCK:END -->