sPhinX: Sample Efficient Multilingual Instruction Fine-Tuning Through N-shot Guided Prompting

TL;DR

This paper introduces sPhinX, a method for improving multilingual instruction fine-tuning of LLMs by creating diverse synthetic datasets and employing N-shot guided prompting, significantly boosting performance across multiple languages.

Contribution

The paper presents a novel multilingual synthetic dataset construction method and a guided fine-tuning strategy that enhance multilingual LLM capabilities with minimal forgetting.

Findings

Improves Mistral-7B performance by 39.8% on multilingual benchmarks.

Enhances Phi-3-Small performance by 11.2%.

Maintains strong English performance with minimal catastrophic forgetting.

Abstract

Despite the remarkable success of large language models (LLMs) in English, a significant performance gap remains in non-English languages. To address this, we introduce a novel approach for strategically constructing a multilingual synthetic instruction tuning dataset, sPhinX. Unlike prior methods that directly translate fixed instruction-response pairs, sPhinX enhances diversity by selectively augmenting English instruction-response pairs with multilingual translations. Additionally, we propose LANGIT, a novel N-shot guided fine-tuning strategy, which further enhances model performance by incorporating contextually relevant examples in each training sample. Our ablation study shows that our approach enhances the multilingual capabilities of Mistral-7B and Phi-3-Small improving performance by an average of 39.8% and 11.2%, respectively, across multilingual benchmarks in reasoning, question answering, reading comprehension, and machine translation. Moreover, sPhinX maintains strong performance on English LLM benchmarks while exhibiting minimal to no catastrophic forgetting, even when trained on 51 languages.