VinaLLaMA-7B

VinaLLaMA-7B is an open-weight Vietnamese large language model built on LLaMA-2 with an additional 800 billion tokens of continued pretraining, described in arXiv 2312.11011 (December 2023). The authors aimed for a truly indigenous Vietnamese model with cultural understanding, not just translated capability. A chat variant, VinaLLaMA-7B-chat, trained on 1M synthetic samples, reported state-of-the-art results on VLSP, VMLU, and Vietnamese Vicuna benchmarks.

A 2.7B variant also exists. As a LLaMA-2 derivative, downstream users should check the LLaMA 2 Community License terms. Benchmark claims are author-reported.