toroe/Qwen3-4B-Instruct-DE-Science-Thinking
toroe/Qwen3-4B-Instruct-DE-Science-Thinking is a 4 billion parameter instruction-tuned causal language model, fine-tuned from Qwen/Qwen3-4B-Instruct-2507. Optimized on the German split of the Nemotron-Multilingual-Reasoning dataset, it excels in German instruction following, step-by-step reasoning, and long-context conversational performance. This model is specifically designed for German-language AI assistants and complex reasoning tasks.
Loading preview...
Overview
toroe/Qwen3-4B-Instruct-DE-Science-Thinking is a 4 billion parameter supervised fine-tuned (SFT) model based on Qwen/Qwen3-4B-Instruct-2507. It was specifically trained on the German (de) split of the DGurgurov/Nemotron-Multilingual-Reasoning dataset to enhance its capabilities in German-language tasks. The training focused on improving instruction following, step-by-step reasoning, and long-context conversational performance in German.
Key Capabilities
- German Instruction Following: Improved ability to understand and execute instructions in German.
- Step-by-Step Reasoning: Enhanced performance on tasks requiring logical, structured explanations in German.
- Long-Context Conversations: Better handling of extended dialogues and document-based interactions, trained with a context length of 16,384 tokens.
- Chat Format: Optimized for chat-based interactions using
apply_chat_template()for best performance.
Intended Uses
- German Assistants and Chatbots: Ideal for building conversational AI systems in German.
- German Reasoning Tasks: Suitable for applications requiring logical problem-solving, math, or structured explanations in German.
- Long-Context Document QA: Effective for question-answering over lengthy German texts.
Limitations
Users should be aware that the model may hallucinate facts, and reasoning is not guaranteed to be correct. Its improvements are primarily for German, and performance near the 16k context limit can depend on prompt structure. It is not aligned for safety-critical deployments and inherits biases from its base model and training data.