nvidia / riva-translate-4b-instruct

Model Overview

Description:

The Riva-Translate-4B-Instruct Neural Machine Translation model translates text in 12 languages. The supported languages are: English(en), German(de), European Spanish(es-ES), LATAM Spanish(es-US), France(fr), Brazillian Portugese(pt-BR), Russian(ru), Simplified Chinese(zh-CN), Traditional Chinese(zh-TW), Japanese(ja),Korean(ko), Arabic(ar).
This model is ready for commercial use.

License/Terms of Use

NIM Package: NVIDIA AI Foundation Models Community License Agreement

Downloadable NIM: NVIDIA AI Foundation Models Community License Agreement

HuggingFace Model: NVIDIA Open Model License Agreement

Model preview in API catalog: NVIDIA Open Model License Agreement

Deployment Geography:

Global

Use Case:

Translators, marketers, and web developers who deliver content in multiple languages.

Release Date:

Huggingface 06/05/2025 via https://huggingface.co/nvidia/Riva-Translate-4B-Instruct

References(s):

[1] Vaswani, Ashish, et al. "Attention is all you need." arXiv preprint arXiv:1706.03762 (2017).
[2] https://github.com/openai/tiktoken
[3] https://en.wikipedia.org/wiki/BLEU
[4] https://github.com/mjpost/sacreBLEU
[5] https://github.com/Unbabel/COMET
[6] NVIDIA NeMo Toolkit

Model Architecture:

Architecture Type: Transformer

Network Architecture: Decoder-only

This model was developed based on Transformer architecture originally presented in "Attention Is All You Need" paper [1]. It is a fine-tuned version of a 4B Base model that was pruned and distilled from nvidia/Mistral-NeMo-Minitron-8B-Base using our LLM compression technique. The model was trained using a multi-stage CPT and SFT. It uses tiktoken [2] as the tokenizer. The model supports a context length of 8K tokens.

Input:

Input Type(s): Text

Input Format: String

Input Parameters: One-Dimensional (1D)

Other Properties Related to Input: This model supports a context length of 8K.

Output:

Output Type(s): Text

Output Format: String

Output Parameters: One-Dimensional (1D)

Other Properties Related to Output: This model supports a context length of 8K.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Prompt Format:

We recommend using the following prompt template, which was used to fine-tune the model. The model may not perform optimally without it.

<s>System
You are an expert at translating text from {Source_language} to {Target_language}.</s>
<s>User
What is the {Target_language} translation of the sentence: {Input_Sentence}?</s>
<s>Assistant\n
<br>

Performance:

COMET score of any2en and en2any direction for Flores-101 dataset

LanguageEng -> LanguageLanguage -> Eng
German0.6630.7575
European Spanish0.74750.7317
Latin American Spanish0.74720.7318
French0.8240.8154
Brazil Portuguese0.8940.8466
Russian0.72340.6427
Simplified Chinese0.66090.701
Traditional Chinese0.63190.6745
Japanese0.72630.6664
Korean0.7120.6801
Arabic0.68880.7073

Software Integration:

Runtime Engine(s): NeMo Framework 24.09

Supported Hardware Microarchitecture Compatibility:

  • NVIDIA Ampere
  • NVIDIA Blackwell
  • NVIDIA Hopper
  • NVIDIA Lovelace

Supported Operating System(s):

  • Linux

Model Version(s):

Riva-Translate-4B-Instruct

Training & Evaluation:

Training Dataset:

Data Collection Method by dataset:

  • Hybrid: Human, Synthetic

Labeling Method by dataset:

  • Automated

Properties: This model is trained on open-sourced datasets and synthetic datasets of text parallel corpora generated via back-translation and monolingual datasets. Each entry in the parallel corpus consists of a text in the source language and its translation in the target language. The monolingual datasets contain texts from each of the 12 target language domains. See bias subcard for language distribution.

Evaluation Dataset:

Link: We used Flores101 [1], NTREX-128 [2], FRMT [3https://www.statmt.org/wmt19/translation-task.html], WMT 19 [4], WMT20 [5] to evaluate the model.

Data Collection Method by dataset:

  • Automated

Labeling Method by dataset:

  • Automated

References:

For more information about these datasets, please see the links below.
[1] https://aclanthology.org/2022.tacl-1.30.pdf
[2] https://aclanthology.org/2022.sumeval-1.4.pdf
[3] https://aclanthology.org/2023.tacl-1.39.pdf
[4] https://www.statmt.org/wmt19/translation-task.html
[5] https://www.statmt.org/wmt20/translation-task.html

Inference:

Acceleration Engine: TensorRT-LLM

Test Hardware:

  • A100
  • A10G
  • H100
  • L40S

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards.

Please report security vulnerabilities or NVIDIA AI Concerns here.