EphAsad/Aristaeus

Aristaeus: A Reasoning-Focused 1.5B Model

Aristaeus, developed by Zain Asad (Eph), is a fine-tuned version of Qwen2.5-1.5B-Instruct, designed to enhance structured, step-by-step reasoning. This 1.5 billion parameter model with a 32K context length focuses on producing deliberate and verifiable chain-of-thought outputs across various domains.

Key Capabilities & Training:

• Enhanced Reasoning: Specifically trained to improve performance in mathematics, science, logic, and code-related reasoning tasks.
• Structured Outputs: Prioritizes clear, step-by-step reasoning over simply achieving correct answers, aiming for explainable thought processes.
• Training Data: Fine-tuned on approximately 47,000 examples from datasets like open-thoughts/OpenThoughts3-1.2M and bespokelabs/Bespoke-Stratos-17k, which include long chain-of-thought traces and competitive programming problems.
• Performance Insights: Manual evaluation against the base model showed Aristaeus winning 3 out of 6 reasoning tasks, drawing 2, and losing 1. It excels in multi-step word problems and spatial constraint reasoning.

Limitations & Future Plans:

• Recursive Tracing: Exhibits limitations in complex recursive call stack tracing, a known capacity ceiling for models of this size.
• Logical Overconfidence: Can sometimes display logical overconfidence, stating conclusions beyond the provided premises due to training for assertive responses.
• Roadmap: Stage 2 plans include fine-tuning for agentic tool use, teaching the model how and when to leverage external tools.

Good for:

• Applications requiring verifiable, step-by-step reasoning in technical domains.
• Use cases where explainability of the thought process is as important as the final answer.
• Developers looking for a compact model with a strong foundation in logical problem-solving.