README.md

metadata

dataset_info:
  features:
    - name: image_id
      dtype: int64
    - name: image
      dtype: image
    - name: width
      dtype: int32
    - name: height
      dtype: int32
    - name: objects
      struct:
        - name: id
          list: int64
        - name: area
          list: int64
        - name: bbox
          list:
            list: float32
            length: 4
        - name: category
          list:
            class_label:
              names:
                '0': Coverall
                '1': Face_Shield
                '2': Gloves
                '3': Goggles
                '4': Mask
    - name: detections
      list:
        - name: bbox
          list: int64
        - name: label
          dtype: string
        - name: sub_label
          dtype: string
        - name: alt_labels
          list: string
    - name: raw_response
      dtype: string
    - name: inference_info
      dtype: string
  splits:
    - name: test
      num_bytes: 3048948
      num_examples: 20
  download_size: 3042132
  dataset_size: 3048948
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
tags:
  - hf-jobs
  - object-detection
  - uv-script

Class-list detection with WeVisOne-VL

Boxes for Coverall, Face_Shield, Gloves, Goggles, Mask on rishitdagli/cppe-5 (split test), from tencent/WeVisOne-VL-4B. 20 images, 20 with at least one detection, 130 detections. 14 boxes share their region with a box of another class (see alt_labels).

These are zero-shot labels: no human checked them. Treat them as a first pass for review.

Columns added

• detections: list of {bbox: [x1, y1, x2, y2], label, sub_label, alt_labels}. bbox is in original-image pixels. sub_label is always empty (WeVisOne gives class labels only); it is there so the schema matches qwen3vl-detect.py. alt_labels lists other classes the model gave the same region (IoU >= 0.8); WeVisOne answers each class on its own, so one object can carry several class names.
• raw_response: the model's raw answer, for re-parsing.
• inference_info: JSON with the model, prompt, classes, image size and settings.

Reproduction

Produced on Hugging Face Jobs (a10g-small) with the wevisone-detect.py recipe from uv-scripts. Run it yourself:

hf jobs uv run --flavor a10g-small https://huggingface.co/datasets/uv-scripts/vlm-object-detection/raw/main/wevisone-detect.py \
    rishitdagli/cppe-5 <output-dataset> --classes Coverall,Face_Shield,Gloves,Goggles,Mask --split test --max-samples 20 --shuffle --seed 42