Model Card for Veps - Russian version 1.0

A model of translation from Vepsian into Russian. In archive initial weights of the model trained with OpenNMT-py (Locomotive). The model has 457M parameters and is trained from scratch. Also presented are model weights converted for Ctranslate2 and a package for installation and use with Argostranslate/Libretranslate.

Model Architecture and Objective

dec_layers: 20
decoder_type: transformer
enc_layers: 20
encoder_type: transformer
heads: 8
hidden_size: 512
max_relative_positions: 20
model_dtype: fp16
pos_ffn_activation_fn: gated-gelu
position_encoding: false
share_decoder_embeddings: true
share_embeddings: true
share_vocab: true
src_vocab_size: 32000
tgt_vocab_size: 32000
transformer_ff: 6144
word_vec_size: 512

How to Use

Using the Model with OpenNMT-py

To fine-tune the Vepsian to Russian translation model using OpenNMT-py, you can modify and use the example configuration file from this repository - config.yml.

Using the Model with LibreTranslate and Argos Translate

To use the Vepsian to Russian translation model with LibreTranslate and Argos Translate, follow these steps:

Download the Model Archive: Ensure you have the translate-vep_ru-1_0.argosmodel file.
Locate the Packages Folder:
- On Linux/MacOS: ~/.local/share/argos-translate/packages
- On Windows: %userprofile%\.local\share\argos-translate\packages
Create the Language Pair Folder:
- Create a folder named vep_ru in the packages directory. If it already exists, delete or move it.
Extract the Model Archive:
- Change the extension of the .argosmodel file to .zip.
- Extract the contents of the .zip file into the vep_ru folder.
Restart LibreTranslate:
- Restart the LibreTranslate application to load the new model.

Citing & Authors

@inproceedings{
title={Model for Veps - Russian translation.},
author={Maksim Migukin, Maksim Kuznetsov, Alexey Kutashov},
year={2024}
}

Credits

Data compiled by Opus.

Includes pretrained models from Stanza.

Data from Vepsian WiKi

Data from Lehme No 2051 // Open corpus of Vepsian and Karelian languages VepKar.

Data from OMAMEDIA

CCMatrix

http://opus.nlpl.eu/CCMatrix-v1.php

If you use the dataset or code, please cite (pdf) and, please, acknowledge OPUS (bib, pdf) as well for this release.

This corpus has been extracted from web crawls using the margin-based bitext mining techniques described here. The original distribution is available from http://data.statmt.org/cc-matrix/

OpenSubtitles

http://opus.nlpl.eu/OpenSubtitles-v2018.php

Please cite the following article if you use any part of the corpus in your own work: P. Lison and J. Tiedemann, 2016, OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation (LREC 2016)

Lynxpda
/

vep_ru