AI4Manufacturing/DD1_1

Roles

Roles: canon repo — annot is the source label, kept machine-parseable as the gold for verification and reward parsing; there is no reasoning column and this repo is not itself a training view. Derived repos (-annotated, -grounding, -region, -mcq) each state their own regime on their own card. Geometry for every record lives in metadata.geometry (below).

DD1 OT VQA Grounding

Answer-only VQA-style grounding data derived deterministically from the OT portion of DD1_cleaned_grounding.

Schema

field	type	meaning
query	string	one of 34 deterministic LPBF OT grounding prompts
image	Image	original 2000×2000 JPEG bytes; never cropped
annot	string	JSON list [{"bbox_xywh":[x,y,w,h]}], or []
reasoning	null	answer-only dataset
cate	string	B
task	string	T-B1
metadata	string	JSON provenance, hashes, boxes and disclosures

Coordinates use native pixels with top-left origin. Boxes are sorted by x, then y. Width and height are derived as xmax-xmin and ymax-ymin.

Counts

• Records: 2667
• Positive images: 1122
• Good/negative images: 1545
• Total boxes: 4339
• Query variants: 34
• Split: train only

Load

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files={"train": "data/train-00000-of-00001.parquet"},
)

annot is the direct SFT answer. reasoning is null on every row.

Reproduce

python3 -m pip install -r requirements.txt
python3 build_dd1_ot_vqa.py \
  --source /path/to/DD1_cleaned_grounding \
  --output /path/to/DD1_OT_VQA_grounding

The build is deterministic and refuses to overwrite an existing output.

Disclosure

Good means no author-annotated overheated region under the source labeling rule; it does not claim absence of every possible manufacturing defect. OT bbox scale varies across layers and may mix local and larger-region annotations.

Geometry (metadata.geometry)

Every record carries a geometry block inside the existing metadata JSON string, so that its gold can be re-derived at any render size. No schema column changed; existing loaders are unaffected.

Coordinates are native pixels of the image in that record (coords_frame: "record_image"). scale is 1.0 throughout — this repo publishes at source resolution, nothing was downscaled at publish time.

"geometry": {
  "image_wh":  [W, H],        // dims of the image in THIS record
  "source_wh": [W, H],        // dims of the original source image
  "scale": 1.0,               // image_wh / source_wh; < 1.0 would disclose a publish-time downscale
  "n_instances": 2,
  "instances": [
    { "instance_id": 1, "bbox_xywh": [x, y, w, h], "min_side_px": 65, "class": null }
  ],
  "n_dropped_subminimum": 0,  // components removed by the filters below
  "union_box_fallback": false,// true => boxes are per-class unions, NOT real instances
  "conventions": { ... }      // see table
}

instances is present even when empty. [] means the record genuinely has no defects; an absent block would mean geometry could not be recovered. Those are different states and are never conflated.

Conventions used to derive it

There is no universal definition of "one defect instance" — it depends on the mask the source shipped. This repo's is stated, not implied:

field	value
algorithm	source_annotation
binarisation	n/a
connectivity	4
merge	none
min_area_px	0
max_instances	None
artifact	fine
fill_floor	None
legibility_floor_px	None
min_side_floor_px	None
spec_sha	5e7f4314c8441019

Provenance and verification

records	2,667
carrying a geometry block	2,667 / 2,667
instances per record	0: 1,545, 1: 77, 2: 107, 3: 455, 4: 202, 5+: 281
total instances	4,339
image dimensions	2000×2000 (2,667)
scale values present	[1.0]

Computed from this repo's own masks and verified against this repo's own published answers before it was written — a recomputation that disagreed with the shipped gold would have aborted the update rather than overwritten it.

⚠ The 16px floor applies at the RENDER, not at native

min_side_px is in native pixels. The model does not see native: Qwen2-VL caps by megapixels AND snaps each dimension to a multiple of 28. So min_side_px >= 16 is the floor tested in the wrong frame. Measured on this repo:

native → rendered (qwen2_vl @ 2.36MP)	2000×2000 → 1512×1512
shipped boxes	4,339
legible at that render (>=16px there)	3,522 (81.2%)

⚠ An earlier version of this section reported the inverse — boxes clearing 16px at native and failing at the render — and that number was misleading. It is frame-relative: publishing at a larger native size lets more boxes clear 16 in the published frame, so more can "fail", which penalises exactly the choice that helps. Measured on 179: publishing native (3024) means a box needs >=32px native to be legible at the render and 86.7% qualify; the previous 1024 publish needed >=47px native and only 69.5% qualified. The native republish improved rendered legibility by 17 points while the old metric scored it as 12.5% "broken". The figure above is the comparable one.

Nothing in the data is frame-dependent — geometry is native and complete. Use forge_model/common/adapt_engine.py, which applies the floor at whatever size the consumer renders.

Using it

Coordinates only stay correct if they are rescaled with the image. A patch-based VLM does not render at native size: Qwen2-VL's processor snaps both dimensions to a multiple of 28, so this repo's 2000×2000 is rendered 1512×1512 and native-pixel boxes are then wrong by a few pixels. forge_model/common/adapt_engine.py regenerates coordinates for a target render size, re-derives counts, and drops records whose gold no longer holds there.

Query text — pooled paraphrases (v2)

Every record's query is drawn from this repository's own query_templates.json (34 prefixes, each followed by the one output directive; the file in this repository is the corrected one), assigned by the build script's hash of the salt and the image's image_sha256 and recorded as metadata.query_template (34 templates in use, top share 3.7%).

v1 drew from the first 34 of these templates with the same hash rule (0 records keep their v1 text). Query wording only: the word 'annotated' was removed from every prefix and from the output directive (the model sees no annotation); the template index per record is unchanged (same salt, same hash of image_sha256), and the repository's query_templates.json is replaced by the corrected one; annot and reasoning (null) are untouched.

Template ↔ gold independence on this build: 2,667 records, 34 templates, worst template p = 0.00964, alpha 2.9e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 2,667 images share one frame size.

Answers, images, masks, split and every other field are byte-identical to v1: this revision was issued from the published parquet itself (tools/requery_published.py), not rebuilt from source, and the pixel-identity guard ran on the embedded images (§8 below).

Image identity — measured at this republish

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	2,667
distinct by decoded pixels	2,667
images carrying more than one record	0
images on both sides of the split	0