EphAsad/Aristaeus

Aristaeus: A Reasoning-Focused Language Model

Aristaeus is a 1.5 billion parameter model developed by EphAsad (Zain Asad), fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its primary goal is to enhance structured, step-by-step reasoning in domains like mathematics, science, logic, and code. This model emphasizes generating deliberate and verifiable chains of thought.

Key Capabilities & Training:

• Reasoning Focus: Trained to improve explicit, step-by-step problem-solving.
• Base Model: Built upon Qwen/Qwen2.5-1.5B-Instruct.
• Training Data: Fine-tuned on approximately 47,000 examples from open-thoughts/OpenThoughts3-1.2M and bespokelabs/Bespoke-Stratos-17k, selected for their reasoning traces in math, science, and competitive programming.
• Evaluation: Manual evaluation against the base model showed Aristaeus winning 3 out of 6 reasoning tasks, drawing 2, and losing 1, particularly excelling in unit conversion, multi-step word problems, and spatial constraint reasoning.

Limitations & Future:

• Known Limitations: Exhibits difficulty with recursive call stack tracing (e.g., Fibonacci sequences) and occasional logical overconfidence, which are common for models of this size and fine-tuning approach.
• Roadmap: Stage 2 plans include fine-tuning for agentic tool use to teach the model how and when to leverage external tools, building upon its reasoning foundation.

Good for:

• Applications requiring verifiable, step-by-step reasoning in technical or logical domains.
• Use cases where explicit chain-of-thought is more critical than raw, unexplainable answers.
• Developers looking for a compact model focused on structured problem-solving.