README.md

metadata

license: mit
language:
  - en
  - ur
pipeline_tag: translation
tags:
  - translation
  - roman-urdu
  - transformer
  - from-scratch

English ↔ Roman Urdu

The trained weights for an 81M-parameter encoder–decoder transformer written from primitives in PyTorch — no nn.Transformer, no AutoModel, no training framework.

Roman Urdu is Urdu written in English letters, which is how Urdu is actually typed in messages. One model translates both directions and detects which one is wanted from the input.

Parameters	81M (d_model 512, 6+6 layers, 8 heads, d_ff 2048)
Vocabulary	24k joint byte-level BPE, shared by both languages
Corpus	1.81M sentence pairs, built for this project
Training	30 epochs; checkpoint chosen on validation loss
BLEU (Tatoeba, held out)	23.2 en→ur, 30.7 ur→en
Weights	fp16 on disk, fp32 at inference

• Code: https://github.com/hasyarshad/roman-urdu-translator
• Demo: https://haseeb-arshad.me/translator
• Space: https://huggingface.co/spaces/hasyarshad/roman-urdu-translator

Use

pip install torch tokenizers huggingface_hub
git clone https://github.com/hasyarshad/roman-urdu-translator.git
cd roman-urdu-translator
python -c "from huggingface_hub import snapshot_download; \
snapshot_download('hasyarshad/roman-urdu-translator', local_dir='export')"
python serve.py "wo larki pyari ha"

Files

File	What it is
model.pt	fp16 weights plus the config needed to rebuild the architecture
tokenizer_joint.json	the joint byte-level BPE tokenizer
roman_urdu_map.json	transliteration map, so Arabic-script input is romanised first
export_info.json	source checkpoint, epoch, validation loss, parameter count