maxzt/Roswaal-8B
Roswaal-8B by maxzt is an 8 billion parameter reasoning model built on a Qwen3-8B base, specifically post-trained using Chain-of-Thought (CoT) distillation. It excels in complex logical and mathematical challenges, achieving 87.64% exact_match on GSM8K, outperforming its base model and many larger alternatives. This model is optimized for deconstructing multi-step problems and performing self-correction, making it highly efficient for demanding analytical tasks.
Loading preview...
Roswaal-8B: A Specialized Reasoning Model
Roswaal-8B, developed by maxzt, is an 8 billion parameter model derived from a deeply uncensored Qwen3-8B base. Its core innovation lies in its post-training via Chain-of-Thought (CoT) distillation, which significantly enhances its reasoning capabilities.
Key Capabilities & Performance
- Dominant Reasoning: Achieves an impressive 87.64% exact_match on the full GSM8K test set (1,319 questions) using 5-shot evaluation, surpassing its base model (Qwen3-8B at 79.4%) and several larger models like Llama 3.1 8B Instruct.
- Advanced CoT: Strictly trained to deconstruct complex prompts, perform step-by-step reasoning, and self-correct within
<think>blocks before providing a final answer. - Efficiency: Developed with high-efficiency training, completed locally on a single NVIDIA RTX 6000 Ada Generation GPU in just over an hour using Unsloth optimization.
Ideal Use Cases
- Complex Problem Solving: Designed for technically demanding, multi-step logical and mathematical challenges.
- Analytical Tasks: Suitable for applications requiring detailed, verifiable reasoning processes.
- Reduced Refusals: Intentionally built to engage with difficult prompts without unnecessary refusals or boilerplate disclaimers.
Roswaal-8B requires a specific generation strategy, utilizing <think>...</think> tags for its internal monologue and recommending low temperature (0.1-0.6) and sufficient max_new_tokens (1024-4096) to allow for thorough reasoning.