EphAsad/AristaeusAgent

AristaeusAgent: Agentic Tool-Use with Explicit Reasoning

EphAsad/AristaeusAgent is a 1.5 billion parameter QLoRA fine-tune by Zain Asad, building upon the EphAsad/Aristaeus model. This model represents Stage 2 of a two-stage training pipeline, specifically designed to integrate structured agentic tool-use capabilities with a "think-before-act" mechanism.

Key Capabilities & Features

• Structured Agentic Tool-Use: Implements a Hermes-style tool-call format, requiring reasoning within <think>...</think> blocks before invoking tools via <tool_call>...</tool_call>.
• Two-Stage Training: Extends the chain-of-thought reasoning established in Stage 1 (Aristaeus) with agentic capabilities, preserving initial reasoning gains.
• Performance Improvement: Achieves a 17.4 percentage point increase in a custom 50-test benchmark compared to its base model, particularly in Reasoning Quality and Multi-Step Planning.
• Tool Refusal Learning: Incorporates negative examples to teach the model when not to call tools, though this remains a primary limitation.
• Open Datasets: Trained exclusively on Apache 2.0 licensed datasets, avoiding API-generated outputs from closed models.

Limitations & Considerations

• Tool Over-triggering: The model still tends to call tools unnecessarily for static knowledge questions, a known limitation partially mitigated by system prompt instructions.
• Hallucination at 1.5B: Due to its parameter count, the model may confabulate supporting details. For production use, a larger base model (e.g., Qwen2.5-3B or 7B) is recommended.
• Recursive Reasoning Failure: Inherits a limitation from its base model regarding deep recursive call stack tracing.

Good for

• Developing Agentic Workflows: Ideal for experimenting with and implementing agentic systems that require explicit reasoning before tool invocation.
• Research into Agent Architectures: Provides a proof-of-concept for a two-stage reasoning-to-agentic pipeline at a smaller scale.
• Tool-Calling Applications: Suitable for tasks where structured tool interaction and clear reasoning steps are crucial.