kyujinpy/Korean-OpenOrca-13B-v2

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

The kyujinpy/Korean-OpenOrca-13B-v2 is a 13 billion parameter auto-regressive language model developed by Kyujin Han, based on the LLaMA2 transformer architecture. It is specifically fine-tuned for Korean language tasks using the OpenOrca-ko-v2 dataset. This model is designed to perform well across various Korean benchmarks, including reasoning, common sense, and question answering.

Loading preview...

Overview

The kyujinpy/Korean-OpenOrca-13B-v2 is an auto-regressive language model developed by Kyujin Han, part of a research consortium involving Media Group Saramgwasup and Marker Inc. Korea. It is built upon the LLaMA2 transformer architecture, specifically using the hyunseoki/ko-en-llama2-13b as its base model. The model was trained using the OpenOrca-ko-v2 dataset, which is a Korean translation of the original OpenOrca dataset, facilitated by DeepL.

Key Capabilities

  • Korean Language Proficiency: Optimized for understanding and generating Korean text.
  • Instruction Following: Fine-tuned on an instruction dataset to improve response quality.
  • Benchmark Performance: Achieves an average score of 48.17 across various Korean benchmarks, including Ko-ARC (43.17), Ko-HellaSwag (54.51), Ko-MMLU (42.90), Ko-TruthfulQA (41.82), and Ko-CommonGen V2 (58.44).

Good for

  • Korean NLP Applications: Ideal for tasks requiring strong Korean language understanding and generation.
  • Research and Development: Suitable for researchers exploring LLaMA2-based models with Korean-specific fine-tuning.
  • Comparative Analysis: Useful for comparing performance against other Korean language models, as benchmark results are provided.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p