VIDraft/Darwin-9B-KOREA
Darwin-9B-KOREA is a 9 billion parameter dual-language model developed by VIDraft, optimized for Korean and English reasoning. This model was created by merging Darwin-9B-Opus and Darwin-9B-NEG using per-layer dynamic linear interpolation, achieving nearly equivalent Korean language capabilities to its 28B-KOREA counterpart at a significantly smaller size. It excels in scenarios requiring English reasoning followed by Korean responses, maintaining high accuracy and Korean output rates. The model supports a context length of 32768 tokens and is recommended for tasks prioritizing accurate answers.
Loading preview...
Overview
VIDraft's Darwin-9B-KOREA is a 9 billion parameter model specifically optimized for dual-language (Korean/English) reasoning. It is a lightweight sibling to the 28B-KOREA model, achieving comparable Korean language performance at approximately one-third the size. This model was created by merging two parent models, Darwin-9B-Opus and Darwin-9B-NEG, using a per-layer dynamic linear interpolation algorithm.
Key Capabilities
- Dual-Language Reasoning: Optimized for English reasoning and generating Korean responses.
- Efficient Performance: Achieves high accuracy (93.1%) and Korean output rates (91.3%) comparable to its larger 28B counterpart, despite its smaller size.
- Parent Model Heritage: Inherits strengths from Darwin-9B-Opus (English reasoning, Korean answers) and Darwin-9B-NEG (negative SFT, synthesis).
- Context Length: Supports a substantial context window of 32768 tokens.
Use Cases
- Accuracy-Focused Tasks: Recommended for applications where answer accuracy is prioritized.
- Chain-of-Thought Reasoning: Best utilized with
max_tokensof 1024 or more to leverage its chain-of-thought characteristics. - Resource-Efficient Deployment: Requires approximately 20GB of GPU memory (BF16), making it suitable for single high-end GPUs like H100/H200/A100 80GB.