Generative Model for Less-Resourced Language with 1 billion parameters

TL;DR

This paper introduces GaMS 1B, a large generative language model for Slovene, trained by adapting an English model with new tokenization and embedding transfer methods, showing promising results in language tasks.

Contribution

The paper presents a novel approach to developing a large generative model for a low-resource language by continuing pretraining of an English model and using specialized embedding transfer techniques.

Findings

GaMS 1B performs well on sentence simplification tasks.

The model lags behind fine-tuned Slovene BERT models in classification.

GaMS achieves comparable or better results than GPT-3.5-Turbo in sentence simplification.

Abstract

Large language models (LLMs) are a basic infrastructure for modern natural language processing. Many commercial and open-source LLMs exist for English, e.g., ChatGPT, Llama, Falcon, and Mistral. As these models are trained on mostly English texts, their fluency and knowledge of low-resource languages and societies are superficial. We present the development of large generative language models for a less-resourced language. GaMS 1B - Generative Model for Slovene with 1 billion parameters was created by continuing pretraining of the existing English OPT model. We developed a new tokenizer adapted to Slovene, Croatian, and English languages and used embedding initialization methods FOCUS and WECHSEL to transfer the embeddings from the English OPT model. We evaluate our models on several classification datasets from the Slovene suite of benchmarks and generative sentence simplification task SENTA. We only used a few-shot in-context learning of our models, which are not yet instruction-tuned. For classification tasks, in this mode, the generative models lag behind the existing Slovene BERT-type models fine-tuned for specific tasks. On a sentence simplification task, the GaMS models achieve comparable or better performance than the GPT-3.5-Turbo model.