EphAsad/AristaeusAgent

AristaeusAgent: Two-Stage Agentic Tool-Use Model

AristaeusAgent is a QLoRA fine-tune of EphAsad/Aristaeus, built upon Qwen2.5-1.5B-Instruct. It represents Stage 2 of a two-stage training pipeline, focusing on adding structured agentic tool-use capabilities to the reasoning foundation established in Stage 1 (Aristaeus).

Key Capabilities

• Agentic Tool-Calling: Implements a <think>...</think> before <tool_call>...</tool_call> behavior, enabling deliberate reasoning prior to tool invocation.
• Hermes-style Format: Utilizes a specific JSON-based tool-call format wrapped in <tool_call> tags, distinct from the base model's raw JSON output.
• Enhanced Reasoning & Planning: Achieved a +17.4 percentage point improvement in overall benchmark score compared to its base, with significant gains in Reasoning Quality (+20 points) and Multi-Step Planning (+14 points).
• QLoRA Fine-tuning: Employs QLoRA (r=16, alpha=32) to preserve Stage 1's chain-of-thought reasoning while extending it with agentic capabilities.

Good for

• Proof-of-Concept for Agentic Pipelines: Demonstrates the viability of a two-stage reasoning-to-agentic pipeline at 1.5B parameters using open datasets.
• Tasks Requiring Deliberate Tool Use: Ideal for scenarios where explicit reasoning and structured tool invocation are critical.
• Research into Agentic LLMs: Provides a foundation for exploring agentic behaviors and tool-calling mechanisms in smaller models.

Limitations

• Tool Over-triggering: The primary limitation is a tendency to call tools unnecessarily for static knowledge questions, scoring lower on Tool Refusal than the base model. This can be partially mitigated with explicit system prompt instructions.
• Hallucination at 1.5B: Due to its parameter size, 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 Stage 1 where deep recursive call stacks can cause the model to lose context.