---
language:
- en
tags:
- computer-vision
- visual-question-answering
- object-detection
- manufacturing
- defect-detection
task_categories:
- visual-question-answering
configs:
- config_name: OT
  data_files: data/OT/*.parquet
- config_name: PB
  data_files: data/PB/*.parquet
---

# LPBF defect-detection VQA

VQA-formatted version of the provided laser powder bed fusion defect dataset.
The `OT` and `PB` configurations are independent and should be treated as two
separate data sources.

## Record format

Each record has `query`, `image`, `annot`, `reasoning`, `cate`, `task`, and
`metadata` (plus a stable `id`). `query` is the VQA question and `annot` is a
natural-language answer: it first says whether a defect exists, then gives the
class and location of every defect. The original Pascal VOC boxes are preserved
in `metadata.objects` as pixel `[xmin, ymin, xmax, ymax]` (`xyxy`) values.
`reasoning` is `null` because no reasoning was supplied by the source data.

The VQA query is: *Identify every visible defect in the image and return its
class and bounding box.*

| Configuration | Images | Positive images | Defect label |
| --- | ---: | ---: | --- |
| OT | 2,674 | 1,122 | `overheated` |
| PB | 2,638 | 1,529 | `defects` |

All images are included. Images without a matching XML and XML files with no
objects become `no_defect` VQA records with the answer “No, there are no
visible defects detected in this image.” `metadata.annotation_status` records
whether a negative example came from an empty XML or the no-XML convention.

## Upload-ready layout

```text
hf_vqa/
├── README.md
└── data/
    ├── OT/train-00000-of-00001.parquet
    └── PB/train-00000-of-00001.parquet
```

Each Parquet file embeds image bytes in the `image` column, so the `hf_vqa/`
folder alone is the upload artifact. The raw JSONL and image folders are only
needed to reproduce it. Use `python3 convert_ot_to_vqa.py`,
`python3 convert_pb_to_vqa.py`, then `python3 build_hf_parquet.py`.

To reproduce the Parquet build in a clean Python environment, install
`pip install -r requirements.txt` first. The XML-to-JSONL converters otherwise
use only the Python standard library.