sergiudanstan/motor-nameplate

Motor Nameplate

A small image dataset of electric motor nameplates collected from public Google Images results, intended for tasks such as:

• Training OCR / document-understanding models to extract nameplate text fields (manufacturer, HP, RPM, voltage, frame, etc.)
• Image classification by manufacturer (ABB, Siemens, Baldor-Reliance, WEG, Hyundai, etc.)
• Object detection (locating the plate on the motor body)
• Few-shot learning / evaluation baselines for industrial vision tasks

The dataset is flat: JPEG files at the repo root, no per-class subdirectories, no labels baked into the directory tree. Labels are provided as a sidecar file (metadata.csv) so the canonical way to load the data is via 🤗 Datasets with data_files.

Dataset structure

.
├── 01.jpg               # 597×335, ~30 KB
├── …
├── 047.jpg              # 638×480, ~58 KB
├── …
├── 286.jpg              # total 286 images
└── metadata.csv         # one row per image; see schema below

• Format: JPEG, ~500–640 px on the long edge, ~30–58 KB average per file
• Layout: flat root (CSV-driven, not imagefolder)
• Total size: ~9 MB images + ~150 KB metadata.csv
• Splits: none — define your own train/val/test split

metadata.csv schema

Column	Type	Notes
file_name	str	matches the image filename at repo root
manufacturer_label	short class	canonical short brand as the top/headline brand: Baldor-Reliance, Yaskawa, Marathon, Imperial-Electric, ABB, Siemens, …, or multi_oem for stacked branding
manufacturer_text	str	the verbatim text spelling the company name on the plate
model_or_cat_no	str / null
serial_no	str / null
p_n	str / null	part number
frame	str / null	e.g. 326T, 215T
hp	number / null	horsepower
kw	number / null	kilowatt rating when present
rpm	number / null	full-load RPM
volts	str / null	verbatim (e.g. 230/460, 240)
amps	str / null	verbatim (e.g. 27/13.5, 71)
enclosure	str / null	e.g. TEFC, ODP, TEFV
phase	int / null	1 or 3
hz	int / null	50 or 60
duty	str / null	e.g. CONT, 60-MIN
amb_c	number / null	ambient temperature °C
insul_class	str / null	B, F, H, …
service_factor	number / null
raw_text	str	full OCR transcription, reading order
confidence	float 0–1	model self-reported on manufacturer_label
error	str	non-empty only for failures (e.g. missing_locally)

Loading

With 🤗 Datasets (recommended)

from datasets import load_dataset

ds = load_dataset(
    "sergiudanstan/motor-nameplate",
    data_files="metadata.csv",
    split="train",
)
# ds[i]["file_name"]   -> e.g. "142.jpg"
# ds[i]["manufacturer_label"]
# ds[i]["hp"]
# ds[i]["confidence"]

Pairing the CSV with the images:

from huggingface_hub import snapshot_download
from pathlib import Path

root = Path(snapshot_download(
    repo_id="sergiudanstan/motor-nameplate",
    repo_type="dataset",
))
ds = load_dataset(str(root / "metadata.csv"), split="train")
def add_image(ex):
    ex["image"] = str(root / ex["file_name"])
    return ex
ds = ds.map(add_image)

Manual download

hf download sergiudanstan/motor-nameplate --repo-type dataset

Direct file URLs

https://huggingface.co/datasets/sergiudanstan/motor-nameplate/resolve/main/142.jpg
https://huggingface.co/datasets/sergiudanstan/motor-nameplate/resolve/main/metadata.csv

Labeling

The auto-labels in metadata.csv were generated by qwen3.6:35b (Ollama) on each plate image, prompted with the schema above and constrained to JSON output (thinking disabled, format=json). Confidence is the model's self-rated score for manufacturer_label. Rows with low confidence may have OCR errors; review triage.md for the bottom-decile and correct by hand before training.

Data sources

Images

• Source: Google Images, search query "motor nameplate" (no language restriction, English UI).
• Collection method: scroll the Google Images results page in headless Chromium, extract every thumbnail URL whose rendered size is ≥150 px on the longer edge, then download each thumbnail.

Limitations

• Many plates have scratched, faded, or partially occluded fields; the OCR columns reflect what's legible at thumbnail resolution.
• "Manufacturer" is sometimes ambiguous: a single plate may show both the OEM (e.g. Yaskawa Electric America) and a co-manufacturer (e.g. MFG. BY Marathon Electric Manufacturing Corp.). The manufacturer_label column uses the top-line brand, but manufacturer_text preserves the verbatim string.
• Visually-similar plates (same source, different JPEG encodes) may both appear in the dataset; perceptual-hash dedup is the user's responsibility.
• A small number of plates are unreadable at this resolution; their fields are null.

Citation

@dataset{sergiu_dan_stan_motor_nameplate,
  title  = {Motor Nameplate},
  author = {sergiudanstan},
  year   = {2026},
  url    = {https://huggingface.co/datasets/sergiudanstan/motor-nameplate}
}