Aura-7b: Agentic AI Workflow Model
Aura-7b is a 7.6 billion parameter language model from Featherlabs, specifically fine-tuned for agentic AI workflows. Built upon the Qwen2.5-7B-Instruct base model, it leverages the Featherlabs Agentic v1 dataset, comprising 14.7K multi-turn agentic conversations, to enhance its capabilities in complex task execution.
Key Capabilities
- Tool Use: Supports structured JSON function calling with defined tool schemas.
- Multi-Step Planning: Proficient at decomposing intricate tasks into manageable, executable steps.
- Chain-of-Thought: Utilizes internal reasoning via
<think> tags to plan actions. - Conversation: Maintains coherent and context-aware multi-turn dialogues.
Performance Insights
While specialized, Aura-7b demonstrates strong performance in specific areas. It achieves 77.6% on GSM8K, outperforming Mistral-7B and Gemma-2-9B, and shows an 8.7% improvement on HellaSwag over its base model, indicating stronger commonsense reasoning. Benchmarks like MMLU, ARC, and TruthfulQA show expected trade-offs due to its specialized agentic fine-tuning, as these do not fully capture its primary strengths in tool use and instruction adherence.
Training Details
The model underwent full Supervised Fine-Tuning (SFT) for 5 epochs on the Featherlabs Agentic v1 dataset, which combines data for function calling, identity framing, and chain-of-thought reasoning. The training utilized Unsloth + TRL (SFTTrainer) on AMD MI300X hardware.
Good For
Developers building AI agents that require robust function calling, multi-step task planning, and structured reasoning will find Aura-7b particularly effective. Its design prioritizes instruction adherence and complex workflow management over general knowledge benchmarks.