sontianye/agenticqwen

AgenticQwen is a Python project that implements a research technique for training small AI language models to be capable "agents", meaning models that can use tools, follow multi-step instructions, and handle complex tasks on their own. It is based on an academic paper (arXiv 2604.21590) that proposed a "dual-flywheel" approach to generating training data, but the paper did not release code, so this repo is described as the first open-source implementation. The core idea involves two data-generation pipelines running in parallel. The "reasoning flywheel" takes problems the model fails on, expands them into new variations, injects different personas, and filters for consistent answers. The "agentic flywheel" takes a persona, generates a set of virtual tools, builds out complex task scenarios using a behavior tree, and then has a teacher AI model solve them step-by-step, recording the whole interaction as training data. Those recorded interactions are then used to train small models (ranging from 1.7 billion to 30 billion parameters) via a reinforcement-learning technique called GRPO, where the model is rewarded based on how well it completes structured sub-goals. You would use this if you are an AI researcher or ML engineer who wants to train a small, efficient language model to handle agentic tasks, such as using tools or following multi-step workflows, without needing massive GPU clusters. The project works with any OpenAI-compatible API (DeepSeek, Qwen, and others), is designed to be resumable if interrupted, and includes 22 unit tests that run without an API key. The tech stack is Python, with YAML config files, the verl training framework, and SGLang for inference.