Optimizing Low-Resource Language Model Training: Comprehensive Analysis of Multi-Epoch, Multi-Lingual, and Two-Stage Approaches

TL;DR

This paper systematically analyzes hyperparameter configurations for low-resource language LLM training, revealing optimal strategies for different data scenarios and providing practical guidelines for efficient model development.

Contribution

It offers a comprehensive exploration of multi-epoch, multi-lingual, and two-stage training approaches, identifying optimal setups and stable model scales for low-resource languages.

Findings

Optimal training approach shifts with target data amount

Stable optimal model scale regardless of data size

Power law relationship for validation loss in single-stage training

Abstract

In this paper, we address the challenge of optimizing training setups for Large Language Models (LLMs) of low-resource language with a limited amount of corpus. Existing works adopt multi-epoch, multi-lingual, and two-stage training to utilize the limited target language corpus efficiently. However, there is still a lack of understanding about the optimal hyperparameter setups for combining these three approaches to train LLMs. We exhaustively explore training setups for low-resource language LLM, combining these three approaches, and found the following insights for efficiently reducing the cost of hyperparameter search: (1) As the amount of target language corpus decreases, the optimal training approach shifts from monolingual single-stage training to multi-lingual two-stage training at a compute budget dependent threshold. (2) The optimal model scale remains stable regardless of the amount of target language corpus, allowing the use of the compute-optimal scale of monolingual training. (3) The optimal number of epochs can be extrapolated from smaller-scale experiments to larger scale using our proposed model. Also, we provide evidence that, in single-stage training, the target language validation loss follows a power law with respect to the target language ratio, with an exponent independent of the amount of data, model scale, and language pair.