EphAsad/Atem-v1-1.5B

Overview of Atem v1

Atem v1 is a 1.5 billion parameter reasoning model developed by EphAsad, built on the Qwen2.5-1.5B-Instruct base model. It was fine-tuned using LoRA on a curated dataset of approximately 115,000 examples distilled from multiple frontier teacher models. This initial stage focuses on establishing strong general reasoning capabilities across various domains.

Key Capabilities and Performance

• Enhanced Reasoning: Atem v1 shows significant improvements in reasoning, mathematics, and coding tasks compared to its base model.
• Mathematical Proficiency: Achieves a +30 percentage point improvement on the GSM8K benchmark, scoring 53.0% compared to the base model's 23.0%. This performance is notable, closing approximately 58% of the capability gap with Qwen2.5-7B-Instruct on this task.
• Code and Analytical Skills: Qualitative evaluations indicate stronger performance in code explanation, implementation, debugging, analytical reasoning, and logic identification.
• Efficient Training: Trained for one epoch with LoRA, preserving base model commonsense capabilities with only a minor 2.4% regression on HellaSwag.

Intended Use Cases

Atem v1 is specifically designed for open-ended reasoning tasks that benefit from structured, accurate thinking:

• Code explanation, implementation, and debugging.
• Mathematical problem-solving with detailed working.
• Analytical reasoning and hypothesis evaluation.
• Concept explanation and comparative analysis.
• Logic, argument, and fallacy identification.

Limitations

• No Thinking Traces: As Stage 1, the model does not produce explicit <think> reasoning traces; this capability is planned for Stage 2 (Atem-Wisdom).
• Mathematical Precision: May exhibit arithmetic slips in intermediate steps on complex multi-step calculations, requiring independent verification for high-stakes problems.