Dria-Agent-a-7B: Agentic LLM with Pythonic Function Calling

Dria-Agent-a-7B is a 7 billion parameter large language model from the Dria-Agent-a series, fine-tuned from Qwen/Qwen2.5-Coder-7B-Instruct. This model is specifically designed for agentic applications, distinguishing itself through its innovative Pythonic function calling mechanism.

Key Capabilities & Differentiators

Pythonic Function Calling: Unlike traditional JSON-based methods, Dria-Agent-a-7B uses Python code blocks to interact with tools, allowing for more flexible and powerful agentic behavior.
One-shot Parallel Multiple Function Calls: The model can execute multiple synchronous processes within a single chat turn, enabling complex solutions that would typically require several conversational turns with other function-calling models.
Free-form Reasoning and Actions: It provides reasoning traces in natural language and actions within Python code blocks, mitigating performance loss often associated with rigid output formats.
On-the-fly Complex Solution Generation: The model generates Python programs, supporting custom logic with conditionals and synchronous pipelines, which is a significant advantage over current JSON-based function calling.

Performance Highlights

Evaluated on the Berkeley Function Calling Leaderboard (BFCL), MMLU-Pro, and the custom Dria-Pythonic-Agent-Benchmark (DPAB), Dria-Agent-a-7B demonstrates strong performance in agentic tasks. Notably, it achieves 70.0 on DPAB (Pythonic, Strict) compared to 44.0 for Qwen2.5-Coder-7B-Instruct, highlighting its specialized capabilities. On BFCL, it shows competitive results, often outperforming its base model and sometimes rivaling gpt-4o in specific categories like "Non-Live Parallel AST" (93.50%) and "Non-Live Simple Exec" (93.29%).

Ideal Use Cases

This model is particularly well-suited for developers building intelligent agents that require:

Complex, multi-step task automation.
Dynamic interaction with external tools via Python code.
Advanced reasoning and planning within a single conversational turn.
Applications where custom logic and conditional execution are crucial for agent behavior.