This is the model card of Plume (Parallel Language Model) with a vocabulary size of 32k.

Plume is the first LLM trained for Neural Machine Translation with only parallel Catalan-Centric data from scratch. It is a language model with the same architecture as Gemma 2B. The model is trained for general translation tasks at sentence level. For more information about training, architecture and interpretability of the model check out the paper; "Investigating the translation capabilities of Large Language Models trained on parallel data only". The preprint is available on arXiv.

In recent years, Large Language Models (LLMs) have demonstrated exceptional proficiency across a broad spectrum of Natural Language Processing (NLP) tasks, including Machine Translation. However, previous methodologies predominantly relied on iterative processes such as instruction fine-tuning or continual pre-training, leaving unexplored the challenges of training LLMs solely on parallel data. In this work, we introduce Plume (Parallel Language Model), a collection of three 2B LLMs featuring varying vocabulary sizes (32k, 128k, and 256k) trained exclusively on Catalan-centric parallel examples. These models perform comparable to previous encoder-decoder architectures on 16 supervised translation directions and 56 zero-shot ones.

For more details regarding the model architecture, the dataset and model interpretability take a look at the paper.

The model is proficient in 16 supervised translation directions that include Catalan and is capable of translating in other 56 zero-shot directions as well.

At the time of submission, no measures have been taken to estimate the bias and added toxicity embedded in the model. However, we are aware that our models may be biased since the corpora have been collected using crawling techniques on multiple web sources. We intend to conduct research in these areas in the future, and if completed, this model card will be updated.

For training, the learning rate is warmed up from 1e-7 to a maximum of 3e-4 over the first 2000 steps. We apply a weight decay of 0.1 and a gradient clipping of 1.0. During training, we set an effective batch size of 81,920 tokens per gradient step distributed over 40 NVIDIA H100-64GB GPUs. We use DeepSpeed with full float32 training. We show in the next table the training hyperparameters:

More training details are specified in the paper. Code for training the model and running other experiments can be found in our GitHub repository.

Below are the evaluation results on Flores-200 and NTREX for supervised MT directions. For more details about model evaluation check out the paper.

The Language Technologies Unit from Barcelona Supercomputing Center.

Feel free to write us at with any questions you may have to {javier.garcia1, carlos.escolano, aleix.santsavall, francesca.delucafornaciari, audrey.mash, xixian.liao, maite.melero}@bsc.es

Copyright(c) 2023 by Language Technologies Unit, Barcelona Supercomputing Center.

This work has been promoted and financed by the Government of Catalonia through the Aina project, by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215337, as well as by DeepR3 (TED2021-130295B-C32) founded by MCIN/AEI/10.13039/501100011033 and European Union NextGeneration EU/PRTR.