sappho192/gemma3-multilingual-translator-270m

Gemma3 Multilingual Translator (270M)

A multilingual translation model based on Gemma3-270m, fine-tuned with LoRA and Bias-corrected EMA on 7.4M translation pairs.

Features

• 6-way translation: Korean, English, Japanese (all directions)
• High quality: 60% loss reduction (4.03 → 1.59) with EMA smoothing
• Easy to use: Simple prompt format with language tags
• ONNX included: Pre-converted ONNX models for PyTorch-free deployment
• Multiple precisions: fp32 and q4 quantized ONNX models included

Quick Start

Using Transformers (PyTorch)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "sappho192/gemma3-multilingual-translator-270m",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("sappho192/gemma3-multilingual-translator-270m")

# Translate Korean to English
prompt = "<src:ko><tgt:en>\n안녕하세요, 만나서 반갑습니다.\n###\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        pad_token_id=tokenizer.pad_token_id
    )

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
translation = result.split("###")[-1].strip()
print(translation)  # "Hello, nice to meet you."

Using ONNX (No PyTorch Required)

from torch_free.inference import TranslatorInferencer

# Load ONNX model (q4 recommended for speed/size)
translator = TranslatorInferencer(
    "sappho192/gemma3-multilingual-translator-270m",
    precision="q4"
)

# Translate
result = translator.translate(
    "안녕하세요, 만나서 반갑습니다.",
    src_lang="ko",
    tgt_lang="en"
)
print(result)  # "Hello, nice to meet you."

Translation Format

<src:SOURCE_LANG><tgt:TARGET_LANG>
SOURCE_TEXT
###

Language codes: ko (Korean), en (English), ja (Japanese)

Examples

Direction	Input	Output
KO→EN	안녕하세요, 만나서 반갑습니다.	Hello, nice to meet you.
EN→KO	Hello, nice to meet you.	안녕하세요, 만나서 반갑습니다.
JA→KO	こんにちは、お元気ですか？	안녕하세요, 건강하세요?
KO→JA	오늘 날씨가 정말 좋네요.	今日の天気は本当にいいですね。

Model Details

Training Specifications

Parameter	Value
Base Model	google/gemma-3-270m
Training Data	7.4M translation pairs
LoRA Rank	16
LoRA Alpha	32
EMA Decay	0.999
Final Loss	1.59

Model Architecture

Parameter	Value
Hidden Size	640
Intermediate Size	2048
Num Layers	18
Num Attention Heads	4
Num KV Heads	1
Head Dim	256
Vocab Size	262,144
Max Position Embeddings	32,768

Included Files

File	Description
model.safetensors	Merged model weights (LoRA + base)
onnx/model.onnx	ONNX model (fp32)
onnx/model_q4.onnx	ONNX model (int4 quantized)
torch_free/	PyTorch-free inference code

ONNX Model Sizes

Precision	Model Size	Avg. Latency
fp32	1.1 GB	0.26s
q4	764 MB	0.17s

Requirements

For PyTorch Inference

• Python 3.11+
• PyTorch 2.0+
• transformers >= 4.55.0

For ONNX Inference (No PyTorch)

• Python 3.11+
• numpy >= 1.24.0
• onnxruntime >= 1.16.0
• tokenizers >= 0.15.0

License

Apache License 2.0

Citation

If you use this model, please cite:

@misc{gemma3-multilingual-translator,
  author = {Taein Kim},
  title = {Gemma3 Multilingual Translator},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/sappho192/gemma3-multilingual-translator-270m}
}