README.md

---
dataset_info:
- config_name: default
  features:
  - name: sample_id
    dtype: string
  - name: image
    dtype: image
  - name: components
    struct:
    - name: arc
      sequence:
        sequence:
          sequence: int16
    - name: bead
      sequence:
        sequence:
          sequence: int16
    - name: groove
      sequence:
        sequence:
          sequence: int16
    - name: nozzle
      sequence:
        sequence:
          sequence: int16
    - name: pool
      sequence:
        sequence:
          sequence: int16
    - name: seam
      sequence:
        sequence:
          sequence: int16
    - name: wire
      sequence:
        sequence:
          sequence: int16
  splits:
  - name: train
    num_bytes: 154779875.0
    num_examples: 2204
  download_size: 154261746
  dataset_size: 154779875.0
- config_name: l1-recognition
  features:
  - name: sample_id
    dtype: string
  - name: image
    dtype: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 93304104.305
    num_examples: 1629
  download_size: 94771619
  dataset_size: 93304104.305
- config_name: l2-grounding
  features:
  - name: sample_id
    dtype: string
  - name: image
    dtype: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 683959330.09
    num_examples: 9141
  download_size: 175395162
  dataset_size: 683959330.09
- config_name: l3-spatial_relations
  features:
  - name: sample_id
    dtype: string
  - name: image
    dtype: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 297585072.8
    num_examples: 4040
  download_size: 146975063
  dataset_size: 297585072.8
- config_name: l4-segmentation
  features:
  - name: sample_id
    dtype: string
  - name: image
    dtype: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 330066212.912
    num_examples: 4556
  download_size: 163462903
  dataset_size: 330066212.912
- config_name: l5-quality_assessment
  features:
  - name: sample_id
    dtype: string
  - name: images
    sequence: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 521838892.28
    num_examples: 2192
  download_size: 220023968
  dataset_size: 521838892.28
- config_name: l6-defect_detection
  features:
  - name: sample_id
    dtype: string
  - name: images
    sequence: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 2150110349.63
    num_examples: 4643
  download_size: 503116815
  dataset_size: 2150110349.63
- config_name: l7-field_operations
  features:
  - name: sample_id
    dtype: string
  - name: images
    sequence: image
  - name: conversations
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  splits:
  - name: train
    num_bytes: 112833914.0
    num_examples: 420
  download_size: 35110179
  dataset_size: 112833914.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
- config_name: l1-recognition
  data_files:
  - split: train
    path: l1-recognition/train-*
- config_name: l2-grounding
  data_files:
  - split: train
    path: l2-grounding/train-*
- config_name: l3-spatial_relations
  data_files:
  - split: train
    path: l3-spatial_relations/train-*
- config_name: l4-segmentation
  data_files:
  - split: train
    path: l4-segmentation/train-*
- config_name: l5-quality_assessment
  data_files:
  - split: train
    path: l5-quality_assessment/train-*
- config_name: l6-defect_detection
  data_files:
  - split: train
    path: l6-defect_detection/train-*
- config_name: l7-field_operations
  data_files:
  - split: train
    path: l7-field_operations/train-*
---

# Weld-Instruct

## Sources and attribution

### l6-defect_detection
- `smaw_*` rows (2,248 rows, 448 tiles): **Ummah et al., "SMAW surface defect dataset"**, Mendeley Data v1 (2026-08-19), doi [10.17632/f7j76vz53p.1](https://doi.org/10.17632/f7j76vz53p.1), licensed **CC BY 4.0**. Only the 448 un-augmented 640×640 tiles are used (author train 358 + test 90); the rotated / flipped / colour-jittered copies are excluded. Boxes are the authors' COCO annotations (spatter, undercut, slag inclusion), converted to [0,1000) coordinates. Multiple-choice explanations were written by a vision-language model from the tile and the annotated boxes and audited by a second model.
- `intel_*` rows: Intel weld-defect photographs (6-way classification, zoomed views).
- numeric-id rows: rikkarth weld-quality grounding data.