OpenRLHF-Agent

OpenRLHF-Agent is a small runtime for running and training tool-using agents with vLLM, SGLang, and OpenAI-compatible APIs.

Install

OpenRLHF-Agent requires Python 3.10 or newer.

git clone https://github.com/Freder-chen/OpenRLHF-Agent.git
cd OpenRLHF-Agent
python -m pip install -e .

Quick Start

Start a compatible vLLM server, then run the math example:

vllm serve Qwen/Qwen3-4B \
  --port 8009 \
  --served-model-name qwen3

python -m examples.math.runtime_demo

Backends

Backend	API
VLLMCompletionBackend	vLLM token-in/token-out API
SGLangCompletionBackend	SGLang native generation API
OpenAIChatBackend	Chat Completions API
OpenAIResponsesBackend	Responses API

Examples

Example	Purpose
examples/math/	Math inference, evaluation, and training
examples/search/	Search-agent inference, evaluation, and training
examples/robot/	Vision-language robot control in LIBERO

See Architecture for component responsibilities and rollout flow.

License

Apache License 2.0.