mt5-small-ea-finetuned

English → Egyptian Arabic (Masri) neural machine translation
Fine-tuned google/mt5-small on a curated English–Egyptian Arabic parallel corpus.

---

Model Summary

Base model	google/mt5-small (~300M parameters)
Task	Text-to-text translation (English → Egyptian Arabic)
Languages	en → arz (Egyptian Arabic, ISO 639-3)
Architecture	Encoder-decoder Transformer (T5)
Training objective	Supervised fine-tuning with teacher forcing
Max sequence length	64 tokens (source & target)
Precision	bf16 / fp16 (auto-detected at training time)

---

Intended Use

✅ Suitable for:

• Translating conversational English into Egyptian Arabic
• Research on low-resource dialectal Arabic MT
• Prototyping dialect-aware downstream applications (chatbots, subtitles, localization)
• Educational / exploratory NLP projects

❌ Not recommended for:

• High-stakes, legal, medical, or safety-critical translation
• Production systems without human verification
• Translating into Modern Standard Arabic (MSA) — use a dedicated MSA model
• Long-form documents (>64 tokens) — chunking required

---

How to Use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "abdullah-khaled/mt5-small-ea-finetuned"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

PREFIX = "translate English to Egyptian Arabic: "
text = "Where are you going tonight?"

inputs = tokenizer(PREFIX + text, return_tensors="pt", max_length=64, truncation=True)
outputs = model.generate(
    **inputs,
    max_length=64,
    num_beams=4,
    early_stopping=True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))