AceGPT & Native Alignment

The leading open Arabic LLM line from Benyou Wang's FreedomIntelligence group (CUHK-Shenzhen) with KAUST/SRIBD partners. AceGPT (2023, arXiv 2309.12053) localized Llama-2 for Arabic via continued pretraining, SFT, and RLAIF with a cultural-sensitivity reward. The anchor here is Native Alignment (NeurIPS 2024, arXiv 2412.03253): moving alignment into pretraining itself — "alignment at pre-training" rather than post-hoc — powering the AceGPT-v2 generation.

All variants (7B–70B across v1/v1.5/v2) are derivatives of Llama/Qwen bases, not from-scratch pretrains.