di-zhang-fdu/openfugu-conductor-3b

OpenFugu Conductor 3B Overview

OpenFugu Conductor 3B is a 3.2 billion parameter model, derived from meta-llama/Llama-3.2-3B-Instruct, developed by di-zhang-fdu. It serves as a core component of the OpenFugu project, which aims to openly reimplement Sakana AI's Fugu-Ultra orchestration line. The model's primary function is to act as a "Conductor," generating detailed agentic workflows that specify which worker performs what action, in what order, and with which prior outputs, given a user request and a worker pool.

Key Capabilities

• Agentic Workflow Generation: Creates structured plans for tool use and task orchestration.
• Tool-Use Planning: Emits a JSON list of tool calls after a brief planning phase (<think>...</think><answer>[json]</answer>).
• GRPO Fine-tuning: Trained using GRPO (Generative Reinforcement Pre-training with Objectives) on the nvidia/ToolScale dataset, which focuses on tool-use and orchestration tasks.
• Verifiable Tool-Call Reward: Training incorporates a reward system that scores the plan's tool-call sequence against ground-truth actions, leading to improved accuracy in tool invocation.

Good For

• Complex Task Automation: Ideal for scenarios requiring multi-step, agentic execution involving various tools.
• Orchestration Systems: Suitable for integrating into systems that need to dynamically plan and execute actions based on user input and available resources.
• Research in Agentic AI: Provides a strong baseline for exploring and developing advanced agentic capabilities, particularly in tool-use and planning.