LLaMA-Pro-Ko-8B Model Card

Model Description

LLaMA-Pro is an advanced iteration of the original LLaMA model, augmented with additional Transformer blocks. Unlike its predecessor, Llama-pro, which was specialized for programming and mathematics, Llama-Pro-Ko is tailored to the language domain, undergoing post-training for enhanced performance.

Development and Training

The NLP & AI Lab at Korea University developed LLaMA-Pro-Ko, a model boasting 8 billion parameters. This model extends LLaMA2-7B by incorporating Korean tokens via vocabulary extension and was further refined by training on a Korean corpus of 10 billion tokens, exclusively without the inclusion of English data.

Language Specialization and Transfer

While previous models like Llama-ko and Llama-2-ko experienced diminished English capabilities as they learned Korean, Llama-Pro's language transfer approach aims to bolster Korean language performance with minimal impact on its English proficiency.

Bilingual Performance Evaluation

LLaMA-Pro-Ko's performance is evaluated on two fronts: its proficiency in English and its mastery of Korean, showcasing its capabilities as a bilingual model.

Korean Evaluation

Open Ko LLM Benchmark

	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2	AVG
Llama-2-7b	31.91	41.68	34.11	48.49	30.34	37.31
beomi/open-llama-2-ko-7b	40.02	50.27	27.60	38.67	42.15	39.74
llama-pro-ko-8b	40.19	51.26	36.80	40.24	43.8	42.46

English Evaluation

Open LLM Benchmark

	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	AVG	diff
meta-llama/Llama-2-7b	53.07	78.59	46.87	38.76	74.03	58.26	0
beomi/llama-2-ko-7b	48.46	75.28	39.56	34.49	72.14	53.99	-4.28
beomi/open-llama-2-ko-7b	46.84	69.48	29.86	35.35	66.30	49.57	-8.70
llama-pro-ko-8b	53.24	77.93	47.06	38.32	72.22	57.75	-0.51