Model Overview
lightblue/DeepSeek-R1-Distill-Qwen-14B-Multilingual is a 14.8 billion parameter model developed by Lightblue, building upon the deepseek-ai/DeepSeek-R1-Distill-Qwen-14B architecture. Its core innovation lies in its multilingual Chain-of-Thought (CoT) fine-tuning, which allows the model to perform both its internal reasoning process ("think") and its final response in the language of the prompt. This contrasts with the original R1 models, which often default to English or Chinese for internal thought processes.
Key Capabilities & Features
- Multilingual CoT Reasoning: Processes and generates responses in the user's input language, improving explainability and relevance for diverse linguistic contexts.
- Enhanced Understandability: Aims to make AI outputs more accessible and transparent for a wider, non-English/Chinese speaking audience.
- Broad Language Support: Demonstrates reliable performance across numerous languages, including higher-resource languages like Japanese, German, and English, as well as various lower-resource languages.
Performance & Usage
Initial evaluations indicate strong performance in generating correct answers and maintaining the specified language for many languages. While higher-resource languages show more consistent accuracy, the model generally performs well across a broad spectrum. For optimal results, a sampling temperature between 0.5-0.7 is recommended, consistent with the original distilled R1 models. Users may also consider setting repetition_penalty to 1.1 or higher for niche languages to mitigate repetition issues. The model was trained on the lightblue/reasoning-multilingual-R1-Llama-70B-train dataset.