br1-pist/NVIDA-Nemotron-3-30B-Nano-AgentCoder
br1-pist/NVIDA-Nemotron-3-30B-Nano-AgentCoder is a 31.6 billion parameter model developed by Bruno Pistone, post-trained from NVIDIA's Nemotron-3-Nano-30B-A3B. It is specifically optimized for agentic behavior in multi-turn, tool-calling conversations, utilizing Group Relative Policy Optimization (GRPO) and Reinforcement Learning from AI Feedback (RLAIF). This model excels at deciding between answering, calling tools, and reacting to tool output in complex interactive scenarios, while preserving the base model's general reasoning capabilities.
Loading preview...
Model Overview
Nemotron-3-Nano-30B-A3B-AgentCoder, developed by Bruno Pistone, is a 31.6 billion parameter model derived from NVIDIA's Nemotron-3-Nano-30B-A3B. Its core innovation lies in its post-training methodology, which employs Group Relative Policy Optimization (GRPO) under Reinforcement Learning from AI Feedback (RLAIF). This approach trains the model on decision points within conversations, rather than entire dialogues, using an LLM judge to score candidate continuations against a rubric. This method significantly improves the model's ability to handle multi-turn agent trajectories and tool calling.
Key Capabilities & Training Highlights
- Agentic Behavior: Optimized for multi-turn agent loops, enabling the model to effectively decide between answering, calling tools, and reacting to tool outputs.
- Tool Calling: Proficient in tool/function calling within coding, data-management, and shell-execution contexts.
- Enhanced Reasoning: Retains the general reasoning and Q&A capabilities inherited from its base model.
- GRPO/RLAIF Training: Achieved a +50.1% improvement in mean judge reward during training, with a substantial reduction in runaway generations (from 22.8% to 4.4%).
- Parameter-Efficient: Only 0.1435% of the weights were trained (45.4M of 31.6B), preserving the base model's general capabilities.
Use Cases
This model is well-suited for applications requiring sophisticated agentic interactions, particularly where a model needs to make dynamic decisions in conversational flows involving external tools. It is ideal for scenarios demanding precise tool selection and multi-step problem-solving.