WhirlwindAI/Translate-15L

---

Overview

Translate-15L is a lightweight multilingual translation model trained on OPUS100.

Rather than maximizing parameter count, the model focuses on providing practical multilingual translation with fast inference and a compact footprint.

It supports translation between English and 14 additional languages in both directions.

---

Quick Start

from transformers import AutoTokenizer, T5ForConditionalGeneration

model = T5ForConditionalGeneration.from_pretrained(
    "WhirlwindAI/Translate-15L"
)

tokenizer = AutoTokenizer.from_pretrained(
    "WhirlwindAI/Translate-15L"
)

text = "Hello, how are you?"

prompt = f"translate en to fr: {text}"

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=50,
    num_beams=4
)

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

---

Supported Languages

Code	Language
🇬🇧 en	English
🇪🇸 es	Spanish
🇫🇷 fr	French
🇩🇪 de	German
🇮🇹 it	Italian
🇵🇹 pt	Portuguese
🇳🇱 nl	Dutch
🇷🇺 ru	Russian
🇨🇳 zh	Chinese
🇯🇵 ja	Japanese
🇰🇷 ko	Korean
🇸🇦 ar	Arabic
🇮🇳 hi	Hindi
🇹🇷 tr	Turkish
🇻🇳 vi	Vietnamese

---

Performance

BLEU Evaluation

Language Overview

---

Highlights

Direction	BLEU
🇬🇧 → 🇫🇷	32.43
🇩🇪 → 🇬🇧	16.93
🇬🇧 → 🇪🇸	6.51
🇫🇷 → 🇬🇧	5.13
🇬🇧 → 🇵🇹	5.05
🇬🇧 → 🇮🇹	4.15

---

Speed

⚡ ~8,151 Tokens / Second

Fast enough for lightweight multilingual applications while remaining compact.

See speed.txt for the complete benchmark.

---

Examples

---

Notes

• English ↔ 14 Languages
• Optimized for compact deployment
• Trained on OPUS100
• Best performance on high-resource European languages
• Performance on low-resource languages remains limited

---

WhirlwindAI

Efficient AI • Practical Research • Open Models