nlpai-lab/KULLM3

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:10.7BQuant:FP8Ctx Length:4kPublished:Apr 2, 2024License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

KULLM3 is a 10.7 billion parameter instruction-following language model developed by NLP&AI Lab, fine-tuned from upstage/SOLAR-10.7B-v1.0. It excels in Korean-speaking instruction-following and chat abilities, closely matching the performance of gpt-3.5-turbo in this domain. The model supports both Korean and English, making it suitable for applications requiring robust bilingual conversational AI.

Loading preview...

KULLM3: A Korean-Centric Instruction-Following LLM

KULLM3, developed by NLP&AI Lab, is a 10.7 billion parameter language model specifically designed for advanced instruction-following and fluent chat in Korean. Fine-tuned from upstage/SOLAR-10.7B-v1.0, it demonstrates performance comparable to gpt-3.5-turbo in its ability to understand and respond to instructions.

Key Capabilities

  • Strong Korean Instruction Following: KULLM3 is noted as one of the best publicly available Korean-speaking language models for instruction adherence.
  • Bilingual Support: The model supports both Korean and English languages.
  • Chat-Optimized: It is designed for engaging and fluent conversational interactions.
  • System Prompt Integration: Performance is optimized when used with the specific system prompt it was trained with, which defines its persona as an AI chatbot from Korea University's NLP&AI Lab.

Training Details

KULLM3 was trained on a mix of data, including vicgalle/alpaca-gpt4 and over 66,000 Korean instruction examples (both GPT-generated and hand-crafted). The training procedure involved a fixed system prompt to guide its conversational style and ethical guidelines.

Good For

  • Applications requiring high-quality Korean instruction-following.
  • Building Korean-speaking chatbots and conversational agents.
  • Research and development in Korean natural language processing.

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