The Idea
The Idea
Translate‑25T is a multilingual translation model fine‑tuned from google/mt5‑small on the OPUS‑100 dataset. It supports translation between English and 18+ other languages in both directions, making it a practical tool for research, education, and lightweight multilingual applications. The focus is on balance: delivering solid translation quality while keeping inference fast and the model compact enough to run in resource‑constrained environments.
Quick Start
from transformers import MT5ForConditionalGeneration, AutoTokenizer
model = MT5ForConditionalGeneration.from_pretrained("WhirlwindAI/Translate-25L")
tokenizer = AutoTokenizer.from_pretrained("WhirlwindAI/Translate-25L")
def translate(text, src, tgt):
prompt = f"translate {src} to {tgt}: {text}"
inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
outputs = model.generate(**inputs, max_new_tokens=128, num_beams=4)
return tokenizer.decode(outputs[0], skip_special_tokens=True)
print(translate("Hello, how are you?", "en", "fr"))
Supported Languages
Supported Languages
Translate‑25T covers a diverse set of languages, including many high‑resource European languages, as well as several Asian and Middle Eastern languages.
| Code | Language | Code | Language |
|---|---|---|---|
| 🇬🇧 en | English | 🇩🇪 de | German |
| 🇫🇷 fr | French | 🇪🇸 es | Spanish |
| 🇷🇺 ru | Russian | 🇨🇳 zh | Chinese |
| 🇯🇵 ja | Japanese | 🇰🇷 ko | Korean |
| 🇸🇦 ar | Arabic | 🇮🇹 it | Italian |
| 🇵🇹 pt | Portuguese | 🇳🇱 nl | Dutch |
| 🇮🇳 hi | Hindi | 🇮🇩 id | Indonesian |
| 🇹🇷 tr | Turkish | 🇻🇳 vi | Vietnamese |
| 🇵🇱 pl | Polish | 🇺🇦 uk | Ukrainian |
| 🇷🇴 ro | Romanian | 🇸🇪 sv | Swedish |
Evaluation
We evaluated Translate‑25T on the OPUS‑100 test set using 50 samples per language pair. The results below show BLEU scores, inference speed, and a combined radar view.
BLEU Scores
Inference Speed
Radar Comparison
Sample Translation (French → English)
| Input (French) | Output (English) |
|---|---|
Bonjour, comment allez‑vous aujourd'hui ? | Hello, how are you here now? |
Model Details
| Property | Value |
|---|---|
| Base Model | google/mt5-small |
| Parameters | 300 Million |
| Architecture | mT5 (Encoder‑Decoder) |
| Training Data | OPUS-100 |
| Languages | 18+ |
| Framework | Hugging Face Transformers |
| License | Apache 2.0 |
Highlights
- Multilingual: translate between English and 18+ languages.
- Efficient: ~300M parameters – lightweight enough for many production scenarios.
- Fast: average inference speed of 52.2 tokens/second on a T4 GPU.
- Research‑friendly: open weights and Apache 2.0 license.
- Practical: trained on a diverse set of parallel sentences from OPUS‑100.
Limitations
- Performance varies across language pairs; high‑resource languages (e.g., English–French) achieve the best BLEU scores.
- The model may struggle with domain‑specific terminology, slang, or very long documents.
- Low‑resource languages are not covered; additional fine‑tuning is recommended for specialised use cases.
Acknowledgements
Built by WhirlwindAI. We thank the Hugging Face team for their ecosystem, and the OPUS project for providing the training data.
License
This model is released under the Apache 2.0 license.
🌪️ WhirlwindAI
Practical AI, built with care.