The DeepHermes Tool Calling Specialist - Atropos RL model is an experimental artifact fine-tuned by Nous Research using our innovative open-source reinforcement learning framework—Atropos. This variant specifically improves the tool calling performance of the DeepHermes 3 Llama-3.1 8B model during its reasoning mode.
Note: This model is intended as an experimental artifact and is not designed for broad, general-purpose use.
Atropos Open Source Framework
Atropos is Nous Research’s open-source Reinforcement Learning environment stack, designed to enhance various aspects of LLM functionalities through structured RL methodologies. We encourage contributions and exploration:
Evaluations on the Berkeley Function Calling benchmark demonstrate significant improvements in tool calling accuracy during reasoning mode, compared to its base model:
Benchmark
Base Accuracy
Atropos RL Accuracy
Improvement
Parallel
0.10
0.46
4.6x
Simple
0.21
0.5175
2.5x
These enhancements are due to RL fine-tuning specifically targeted at improving reasoning-based tool calling capabilities.
Eval set accuracy results:
image/png
Key Features
Improved Tool Calling in Reasoning Mode: Reinforcement learning significantly boosts tool usage during complex reasoning tasks.
Open-Source RL Framework: Utilizes the fully open-source Atropos RL Environments.
Active Open Source Community: Contributions welcomed on the Atropos GitHub.
Upcoming SOTA RL Trainer: A state-of-the-art open-source reinforcement learning trainer by Nous Research is coming soon.
Usage
This model supports multiple inference modes including:
Reasoning (Deep Thinking Mode)
Standard Chat/Instruction Mode
Structured JSON Outputs
Function Calling
Detailed documentation and example inference code are available:
Note: You must first place DeepHermes' reasoning system prompt, and then append your function calling system prompt after for it to do reasoning and tool calling simultaneously.
1@misc{
2 title={DeepHermes Tool Calling Specialist - Atropos RL},
3 author={Teknium and Dakota Mahan and Roger Jin and Chen Guang and Jai Suphavadeeprasit and Jeffrey Quesnelle},
4 year={2025},
5 url={https://huggingface.co/NousResearch/DeepHermes-Tool-Calling-Specialist-Atropos-RL}
6}
Community and Support
For questions, issues, or findings, please open issues or discussions in the respective GitHub repositories:
