Model Overview
The AIDX-ktds/ktdsbaseLM-v0.13-onbased-llama3.1 is an 8 billion parameter language model developed by AIDX-ktds, built upon the Llama3.1 architecture. It has been fine-tuned using the Supervised Fine-Tuning (SFT) method to specialize in the Korean language and cultural contexts. The model incorporates 53 distinct domains of proprietary Korean data, enabling it to understand and reflect Korean societal values and culture.
Key Capabilities
- Text Generation: Produces coherent and contextually relevant Korean text.
- Dialogue Inference: Understands and responds to conversational nuances.
- Document Summarization: Condenses lengthy Korean documents into concise summaries.
- Question Answering (Q&A): Provides accurate answers to questions across various domains.
- Sentiment Analysis: Identifies and interprets emotional tones in Korean text.
- Natural Language Processing (NLP): Supports a wide range of general NLP tasks.
Training Data
The model was trained on a proprietary dataset totaling 3.6GB, comprising 2.33 million Q&A, summarization, and classification data points. This includes 1.33 million multiple-choice questions from 53 domains (e.g., Korean history, society, finance, law, mathematics, biology, physics, chemistry) trained with Chain of Thought, and 1.3 million subjective questions from 38 domains (e.g., Korean history, finance, law, tax, mathematics).
Good For
- Education: Generating explanations and answering questions on subjects like history, math, and science.
- Business: Providing answers to legal, financial, and tax-related queries, and summarizing documents.
- Research & Culture: Performing NLP tasks tailored to Korean society and culture, including sentiment analysis and content creation.
- Customer Service: Developing conversational agents capable of generating tailored responses.
Limitations
While highly specialized for Korean language and culture, the model may exhibit reduced accuracy for other languages or cultures due to data scarcity in those areas. It also has limited reasoning capabilities for complex logical problems and may produce biased responses if trained on biased data.