197 lines
4.3 KiB
Markdown
197 lines
4.3 KiB
Markdown
---
|
||
language:
|
||
- en
|
||
- th
|
||
license: apache-2.0
|
||
pipeline_tag: translation
|
||
tags:
|
||
- translation
|
||
- thai
|
||
- english
|
||
- gemma-3
|
||
- typhoon
|
||
- machine-translation
|
||
- bilingual
|
||
- sft
|
||
base_model: typhoon-ai/typhoon2.1-gemma3-4b
|
||
|
||
|
||
---
|
||
|
||
# 🇹🇭🇺🇸 Typhoon 2.1 Gemma 3 KordTranslate EN-TH 4B
|
||
|
||
<p align="center">
|
||
<img src="https://i.ibb.co/q3H1WPgj/logo.png" width="220" alt="KordTranslate Logo">
|
||
</p>
|
||
|
||
**Typhoon 2.1 Gemma 3 KordTranslate EN-TH 4B** is a bilingual English ↔ Thai translation model fine-tuned from **`typhoon-ai/typhoon2.1-gemma3-4b`** for high-quality machine translation between Thai and English.
|
||
|
||
The model is instruction-tuned specifically for translation tasks using bilingual translation datasets while preserving the strong language understanding inherited from the Typhoon base model.
|
||
---
|
||
|
||
# ✨ Features
|
||
|
||
- 🇺🇸 English → Thai translation
|
||
- 🇹🇭 Thai → English translation
|
||
- Instruction-tuned for translation
|
||
- Built on **Typhoon2.1 Gemma 3 4B**
|
||
- Lightweight and suitable for local inference
|
||
|
||
---
|
||
|
||
| Item | Value |
|
||
|------|-------|
|
||
| Model | `KordAI/Typhoon-Gemma3-KordTranslate-EN-TH-4B` |
|
||
| Base Model | `typhoon-ai/typhoon2.1-gemma3-4b` |
|
||
| Architecture | Gemma 3 (4B) |
|
||
| Task | Machine Translation |
|
||
| Languages | English, Thai |
|
||
| Fine-tuning | Supervised Fine-Tuning (SFT) |
|
||
|
||
---
|
||
|
||
This model is optimized for:
|
||
|
||
- English → Thai translation
|
||
- Thai → English translation
|
||
- Localization
|
||
- Educational applications
|
||
- Chatbot translation
|
||
- API integration
|
||
- Batch document translation
|
||
|
||
|
||
For the best performance, use the prompt templates below.
|
||
|
||
---
|
||
|
||
# 📝 Prompt Templates
|
||
|
||
For optimal translation quality, structure your inputs exactly as shown below depending on the translation direction.
|
||
|
||
## English → Thai
|
||
|
||
```text
|
||
Translate the following English text to Thai.
|
||
|
||
### English:
|
||
[Your English text here]
|
||
```
|
||
|
||
## Thai → English
|
||
|
||
```text
|
||
Translate the following Thai text to English.
|
||
|
||
|
||
### Thai:
|
||
[Your Thai text here]
|
||
|
||
|
||
```
|
||
|
||
---
|
||
|
||
# 🐍 Inference Code (Python / Transformers)
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
|
||
# Hugging Face repository
|
||
model_name = "KordAI/Typhoon-Gemma3-KordTranslate-EN-TH-4B"
|
||
|
||
|
||
# Load tokenizer and model
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
|
||
prompt = """
|
||
Translate the following English text to Thai.
|
||
|
||
|
||
### English:
|
||
But friends, I am so gay, that if I had a wife, I would encourage her to cheat on me.
|
||
"""
|
||
|
||
messages = [
|
||
{"role": "user", "content": prompt}
|
||
|
||
]
|
||
|
||
text = tokenizer.apply_chat_template(
|
||
messages,
|
||
tokenize=False,
|
||
add_generation_prompt=True
|
||
|
||
|
||
)
|
||
|
||
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
||
|
||
generated_ids = model.generate(
|
||
**model_inputs,
|
||
max_new_tokens=512
|
||
)
|
||
|
||
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):]
|
||
|
||
translation = tokenizer.decode(
|
||
output_ids,
|
||
skip_special_tokens=True
|
||
).strip()
|
||
|
||
|
||
print(translation)
|
||
```
|
||
|
||
---
|
||
|
||
# 🏋️ Training
|
||
|
||
This model is fine-tuned from **`typhoon-ai/typhoon2.1-gemma3-4b`** using supervised instruction tuning on bilingual English–Thai translation datasets.
|
||
|
||
The model is trained to perform direct translation between:
|
||
|
||
- 🇺🇸 English → Thai
|
||
- 🇹🇭 Thai → English
|
||
|
||
while maintaining natural, fluent, and context-aware outputs.
|
||
|
||
---
|
||
|
||
# ⚠️ Limitations
|
||
|
||
- Translation quality may decrease on highly specialized domains such as medical, legal, or scientific texts.
|
||
- Proper nouns and newly coined terminology may require manual verification.
|
||
- The model may occasionally hallucinate or paraphrase ambiguous inputs.
|
||
- Human review is recommended for production-critical translations.
|
||
|
||
|
||
---
|
||
Special thanks to:
|
||
|
||
- **Typhoon AI** for the excellent **`typhoon2.1-gemma3-4b`** base model.
|
||
- **KordAI** for developing and fine-tuning the translation model.
|
||
- **Unsloth** for enabling efficient supervised fine-tuning.
|
||
- **llama.cpp** for the GGUF format and inference ecosystem.
|
||
- The open-source AI community for advancing multilingual language models.
|
||
|
||
---
|
||
|
||
# 📖 Citation
|
||
|
||
```bibtex
|
||
@misc{kordtranslate2026,
|
||
title={Typhoon Gemma3 KordTranslate EN-TH 4B},
|
||
author={KordAI},
|
||
year={2026},
|
||
publisher={Hugging Face},
|
||
howpublished={https://huggingface.co/KordAI/Typhoon-Gemma3-KordTranslate-EN-TH-4B}
|
||
}
|
||
``` |