sh-024/LFM2.5-1.2B-Instruct-Korean
sh-024/LFM2.5-1.2B-Instruct-Korean is a 1.2 billion parameter instruction-following language model, fine-tuned by sh-024, based on LiquidAI's LFM2.5-1.2B-Instruct. This model specializes in Korean language tasks, including question answering, general conversation, and instruction-following, with a context length of 32768 tokens. It was trained on Korean-centric datasets to generate natural, consistent, and contextually appropriate responses in Korean. The model is optimized for stable generalization performance in Korean conversational and instructional contexts.
Loading preview...
Model Overview
LFM2.5-1.2B-Instruct-Korean is a 1.2 billion parameter instruction-following language model, fine-tuned by sh-024, specifically for Korean language tasks. Built upon LiquidAI/LFM2.5-1.2B-Instruct, this model aims to enhance performance in Korean question answering, general conversation, and instruction-following.
Key Capabilities
- Korean Language Proficiency: Designed to generate natural, consistent, and contextually appropriate responses in Korean.
- Instruction Following: Fine-tuned to accurately follow instructions provided in Korean.
- Conversational AI: Capable of engaging in general conversations in Korean.
- Stable Training: Demonstrated stable generalization performance with a very small gap between training and evaluation loss (1.1046 vs. 1.1015), suggesting no significant overfitting.
Training Details
The model was fine-tuned for one full epoch (25,652 steps) on Korean-centered datasets including maywell/koVast, CarrotAI/ko-instruction-dataset, and MarkrAI/KoCommercial-Dataset. This diverse dataset includes instruction-response pairs, general conversational data, and commercially oriented text to cover a broad range of Korean expressions and contexts.
Good For
- Applications requiring Korean-specific instruction following.
- Developing Korean conversational agents.
- Tasks involving Korean question answering.
Limitations
While training metrics indicate stability, these do not fully reflect real-world performance. Further evaluations are planned for response quality, factuality, and safety.