Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AceGPT, Localizing Large Language Models in Arabic

Huang Huang, Fei Yu, Jianqing Zhu, Xuening Sun, Hao Cheng, Dingjie Song, Zhihong Chen, Abdulmohsen Alharthi, Bang An, Juncai He, Ziche Liu, Zhiyi Zhang, Junying Chen, Jianquan Li, Benyou Wang, Lian Zhang, Ruoyu Sun, Xiang Wan, Haizhou Li, Jinchao Xu

TL;DR

AceGPT tackles the localization gap in Arabic LLMs by assembling a culture-aware pipeline comprising localized pre-training, Arabic instruction-tuning with native GPT-4 outputs, and RLHF using a localized reward model. The approach yields state-of-the-art results among open Arabic LLMs across multiple benchmarks, including ACVA, Arabic MMLU, and Vicuna-80/AlpacaEval, while also introducing ACVA as a focused localization benchmark. Key contributions include establishing the first open-source Arabic LLM pipeline spanning pre-training, SFT, and RLHF, achieving competitive performance against GPT-3.5 Turbo, and providing a comprehensive evaluation framework for Arabic cultural alignment. The work demonstrates that targeted localization data and culturally aware feedback signals can meaningfully enhance Arabic language understanding and generation in alignment with local values and norms.

Abstract

This paper is devoted to the development of a localized Large Language Model (LLM) specifically for Arabic, a language imbued with unique cultural characteristics inadequately addressed by current mainstream models. Significant concerns emerge when addressing cultural sensitivity and local values. To address this, the paper proposes a comprehensive solution that includes further pre-training with Arabic texts, Supervised Fine-Tuning (SFT) utilizing native Arabic instructions, and GPT-4 responses in Arabic, alongside Reinforcement Learning with AI Feedback (RLAIF) employing a reward model attuned to local culture and values. The goal is to cultivate culturally cognizant and value-aligned Arabic LLMs capable of accommodating the diverse, application-specific needs of Arabic-speaking communities. Comprehensive evaluations reveal that the resulting model, dubbed `AceGPT', sets the state-of-the-art standard for open Arabic LLMs across various benchmarks. Codes, data, and models are in https://github.com/FreedomIntelligence/AceGPT.

AceGPT, Localizing Large Language Models in Arabic

TL;DR

AceGPT tackles the localization gap in Arabic LLMs by assembling a culture-aware pipeline comprising localized pre-training, Arabic instruction-tuning with native GPT-4 outputs, and RLHF using a localized reward model. The approach yields state-of-the-art results among open Arabic LLMs across multiple benchmarks, including ACVA, Arabic MMLU, and Vicuna-80/AlpacaEval, while also introducing ACVA as a focused localization benchmark. Key contributions include establishing the first open-source Arabic LLM pipeline spanning pre-training, SFT, and RLHF, achieving competitive performance against GPT-3.5 Turbo, and providing a comprehensive evaluation framework for Arabic cultural alignment. The work demonstrates that targeted localization data and culturally aware feedback signals can meaningfully enhance Arabic language understanding and generation in alignment with local values and norms.

Abstract

This paper is devoted to the development of a localized Large Language Model (LLM) specifically for Arabic, a language imbued with unique cultural characteristics inadequately addressed by current mainstream models. Significant concerns emerge when addressing cultural sensitivity and local values. To address this, the paper proposes a comprehensive solution that includes further pre-training with Arabic texts, Supervised Fine-Tuning (SFT) utilizing native Arabic instructions, and GPT-4 responses in Arabic, alongside Reinforcement Learning with AI Feedback (RLAIF) employing a reward model attuned to local culture and values. The goal is to cultivate culturally cognizant and value-aligned Arabic LLMs capable of accommodating the diverse, application-specific needs of Arabic-speaking communities. Comprehensive evaluations reveal that the resulting model, dubbed `AceGPT', sets the state-of-the-art standard for open Arabic LLMs across various benchmarks. Codes, data, and models are in https://github.com/FreedomIntelligence/AceGPT.
Paper Structure (52 sections, 2 equations, 2 figures, 23 tables)

This paper contains 52 sections, 2 equations, 2 figures, 23 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Evaluation
  4. Contributions
  5. Recipe of AceGPT
  6. Motivation: The Localization Issue
  7. Methodology of AceGPT
  8. Localized Pre-Training
  9. Localized Supervised Fine-Tuning
  10. Reinforcement Learning from AI Feedback
  11. Localization Evaluation
  12. Evaluation Protocol
  13. Experiment Results
  14. Overall Evaluation
  15. Evaluation Protocol
  16. ...and 37 more sections

Figures (2)

  • Figure 1: (a) Correlations between the reward model and GPT-4 and (b) reward distribution.
  • Figure 2: User interface of human annotation. Response positions are randomized to alleviate biases.