EphAsad/Atem-8B

Atem-8B: A Reasoning-Focused 8B Model

Atem-8B, developed by EphAsad, is an 8 billion parameter model built upon Qwen3-8B. It utilizes a unique CoT-preserving single-pass Supervised Fine-Tuning (SFT) approach to distill multi-domain reasoning capabilities from advanced teacher models, while maintaining the base model's inherent thinking abilities. This method avoids the common issue of erasing native reasoning during initial training stages.

Key Capabilities

• Enhanced Reasoning: Demonstrates improved performance in commonsense reasoning tasks like Winogrande (+4.6pp) and HellaSwag (+1.7pp).
• Multi-Domain Expertise: Proficient in mathematical reasoning, code explanation and debugging, scientific explanation, and general analytical tasks.
• CoT-Preserving SFT: Trained with a single pass using 16-bit LoRA on a diverse corpus of 58,980 records, including Chain-of-Thought (CoT) traces.
• Flexible GSM8K Evaluation: While showing a slight regression in strict GSM8K evaluation due to a shift towards \boxed{} answer formatting, a flexible evaluator recovers most of the performance, indicating a formatting artifact rather than a capability loss.

Good For

• Multi-step mathematical reasoning: Ideal for problems requiring detailed, step-by-step calculations and explanations.
• Code analysis and generation: Useful for understanding, implementing, and debugging code.
• Analytical and scientific reasoning: Excels in tasks demanding structured thought and technical depth.
• Commonsense and logical problem-solving: Suited for scenarios where contextual understanding and logical deduction are critical.