README.md

metadata

language:
  - ko
  - en
base_model:
  - meta-llama/Llama-3.2-1B-Instruct

@ 2024.10.07 Model torchtorchkimtorch/Llama-3.2-Korean-GGACHI-1B-Instruct-v1 Released!

@ 2024.10.18 Performance for KOBEST of Llama-3.2-Korean-GGACHI-1B-Instruct-v1 has been updated!

Llama-3.2-Korean-GGACHI-1B-Instruct-v1

모델 설명 (Model Description)

GGACHI-1B-Instruct-v1는 Llama-3.2-1B-Instruct 모델을 기반으로 하는 한국어 태스크 수행에 최적화된 instruction-tuned 언어 모델입니다. 230,000개 이상의 고품질 한국어 데이터셋을 사용하여 fine-tuning되었습니다.

GGACHI-1B-Instruct-v1 is an instruction-tuned language model optimized for Korean language tasks, based on the Llama-3.2-1B-Instruct model. It has been fine-tuned using over 230,000 high-quality Korean language datasets.

모델 성능 (Model Performance)

- 0 shot

Task	Model	Accuracy
kobest_boolq	Llama-3.2-1B-Instruct	0.502
	GGACHI	0.502
kobest_copa	Llama-3.2-1B-Instruct	0.504
	GGACHI	0.521
kobest_hellaswag	Llama-3.2-1B-Instruct	0.358
	GGACHI	0.380
kobest_sentineg	Llama-3.2-1B-Instruct	0.476
	GGACHI	0.594

- 5 shot

Task	Model	Accuracy
kobest_boolq	Llama-3.2-1B-Instruct	0.571
	GGACHI	0.565
kobest_copa	Llama-3.2-1B-Instruct	0.526
	GGACHI	0.549
kobest_hellaswag	Llama-3.2-1B-Instruct	0.364
	GGACHI	0.398
kobest_sentineg	Llama-3.2-1B-Instruct	0.725
	GGACHI	0.795

- 10 shot

Task	Model	Accuracy
kobest_boolq	Llama-3.2-1B-Instruct	0.593
	GGACHI	0.571
kobest_copa	Llama-3.2-1B-Instruct	0.525
	GGACHI	0.549
kobest_hellaswag	Llama-3.2-1B-Instruct	0.356
	GGACHI	0.394
kobest_sentineg	Llama-3.2-1B-Instruct	0.768
	GGACHI	0.821

Contact

• 김민혁(Minhyuk Kim)
  Mail: [email protected]
  LinkedIn : https://www.linkedin.com/in/mhkim0929/