A sentence-aligned English–Urdu parallel corpus of 4,150 segment pairs extracted from ~32 hours of vocational training video content, covering three trade domains: electrician work, carpentry, and plumbing. Human-validated translations are available for 85.2% of segments.

This dataset was created to support domain-adaptive machine translation for vocational education in Urdu — a language spoken by ~230 million people but severely underrepresented in domain-specific NLP resources.

Dataset Details

Developed by: Mujtaba Umar, BSc Data Science, National University of Sciences and Technology (NUST), Pakistan
Language pair: English (eng_Latn) → Urdu (urd_Arab)
Domains: Electrician, Carpentry, Plumbing
License: CC-BY 4.0
Associated model: mujtabammarr/nllb-600M-vocational-urdu

Dataset Structure

Splits

Split	Size	Description
train	3,442	Training set (includes 612 non-validated segments)
validation	354	Development set (human-validated only)
test	354	Held-out test set (human-validated only)
Total	4,150

Columns

Column	Type	Description
`english`	string	Source English text (vocational video transcript segment)
`target`	string	Urdu translation — human-validated where available
`source_file`	string	Name of the originating video
`target_col`	string	`validation` = human-corrected; `urdu` = initial machine-assisted translation

Example

python

{
  "english": "Here we've added a contact on relay R1 in parallel with the switch. This is called a latching circuit.",
  "target": "یہاں ہم نے سوئچ کے متوازی ریلے R1 پر ایک رابطہ شامل کیا ہے۔ اسے لیچنگ سرکٹ کہا جاتا ہے۔",
  "source_file": "63_Cambridge_Elevating_-_Basic_Electrical_Theory_Part_2",
  "target_col": "validation"
}

Source Material

The dataset was derived from ~155 English-language vocational training videos totalling approximately 32 hours. Videos cover three trade domains:

Electrical: Basic circuit theory, AC/DC circuits, wiring, GFCI outlets, panel installation, dimmer switches, ceiling fans, light circuits, subpanel installation
Carpentry: Hand tools, power tools (circular saw, jigsaw, router, miter saw, table saw), joinery, framing, stair construction, door installation, window trim, cabinet making, woodworking terminology
Plumbing: Pipe materials (PEX, copper, PVC, iron), fixture installation, drainage systems, leak repair, shutoff valves, toilet mechanics, backflow preventers

Collection & Annotation

Each video was transcribed to produce English text segmented at natural speech boundaries. Segments were then translated to Urdu in two stages:

Initial translation: Machine-assisted Urdu translation
Human validation: Native Urdu speakers with trade knowledge reviewed and corrected the translations

The target_col column indicates which stage produced the target translation. For training purposes, using only validation rows is recommended for highest quality.

Each segment retains start and end timestamps from the original video, enabling future work on speech translation and audio-visual alignment.

Linguistic Notes

Trade-specific terms without established Urdu equivalents are rendered via transliteration into Urdu script (e.g., سٹرنگرز for "stringers", لیچنگ سرکٹ for "latching circuit", پی ای ایکس for "PEX"). This reflects authentic vocational usage in Pakistan, where English trade terminology is commonly borrowed into spoken Urdu.

Usage

python

from datasets import load_dataset

ds = load_dataset("MujtabaUmar/vocabur-en-ur-vocational")

# Use validated data only
validated = ds.filter(lambda x: x["target_col"] == "validation")

# Example: print first training pair
print(ds["train"][0]["english"])
print(ds["train"][0]["target"])

Benchmark Results

Evaluated on the held-out test split (354 segments) using sacrebleu.

Model	BLEU	chrF++	Δ BLEU
NLLB-200-distilled-600M (zero-shot baseline)	24.22	51.80	—
NLLB-200-distilled-600M (fine-tuned on VocabUR)	40.80	63.12	+16.58

BLEU measures n-gram precision against the reference translation (higher = better, max 100).
chrF++ measures character n-gram F-score — well-suited for morphologically rich languages like Urdu (higher = better, max 100).

Associated Model

A fine-tuned NLLB-200-distilled-600M model trained on this dataset is available at MujtabaUmar/nllb-600M-vocational-urdu.

Intended Use

Fine-tuning or evaluating MT models for vocational English → Urdu translation
Research on domain adaptation for low-resource language pairs
Benchmarking NMT systems on South Asian vocational content
Building subtitling or dubbing pipelines for Urdu-language vocational video

Out-of-Scope Use

General-domain translation benchmarking
Trades not covered (automotive, healthcare, agriculture, etc.)
Languages other than English and Urdu

Limitations

Coverage is limited to three trade domains; models trained on this data may not generalise to other vocational fields
Human validation coverage is 85.2%; remaining 14.8% are initial machine-assisted translations and may contain errors
Single reference translations only; inter-annotator agreement was not measured
Video sources are predominantly North American English, which may affect terminology for Urdu speakers in different regions

Citation

bibtex

@misc{umar2025vocabur,
  title={Domain-Adaptive Neural Machine Translation for Vocational Education in Urdu},
  author={Umar, Mujtaba},
  year={2025},
  institution={National University of Sciences and Technology (NUST)},
  note={VocabUR: English-Urdu Vocational Parallel Corpus}
}

Contact

Mujtaba Umar — BSc Data Science, NUST, Pakistan mumar.bsds24seecs@seecs.edu.pk

mujtabaumarr/nllb-600M-vocational-urdu

VocabUR: English–Urdu Vocational Parallel Corpus