kyujinpy/ko-platypus-kiwi-13B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kLicense:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Warm

The kyujinpy/ko-platypus-kiwi-13B is a 13 billion parameter auto-regressive language model developed by Kyujin Han (kyujinpy) and a consortium including Media Group Saramgwasoop and Marker. Built upon the LLaMA2 transformer architecture, it is fine-tuned for Korean language tasks. This model demonstrates strong performance in Korean benchmarks, particularly excelling in Ko-CommonGen V2, making it suitable for Korean natural language generation and understanding applications.

Loading preview...

KOR-Orca-Platypus-kiwi🥝: A Korean-Optimized LLaMA2 Model

Developed by Kyujin Han (kyujinpy) in collaboration with Media Group Saramgwasoop and Marker, ko-platypus-kiwi-13B is a 13 billion parameter auto-regressive language model. It is based on the robust LLaMA2 transformer architecture, specifically fine-tuned from the hyunseoki/ko-en-llama2-13b base model.

Key Capabilities & Performance

This model is specialized for Korean language processing, leveraging a dedicated training dataset, kyujinpy/KOR-Orca-Platypus-kiwi. It exhibits competitive performance across various Korean benchmarks:

  • Average Score: 48.97
  • Ko-ARC: 42.41
  • Ko-HellaSwag: 54.29
  • Ko-MMLU: 41.98
  • Ko-TruthfulQA: 40.05
  • Ko-CommonGen V2: 66.12 (its strongest area)

Its notable score in Ko-CommonGen V2 suggests particular strength in Korean common sense generation tasks.

Licensing

The model is released under the cc-by-nc-sa-4.0 license, allowing for non-commercial use with attribution and share-alike terms.