chanyoungkim/grpo-qwen2.5-1.5b-alfworld-step130
The chanyoungkim/grpo-qwen2.5-1.5b-alfworld-step130 is a 1.54 billion parameter Qwen2.5-Instruct model fine-tuned using Generative Reinforcement Learning with Policy Optimization (GRPO) on the ALFWorld environment. This model is specifically optimized for agentic tasks within text-based interactive environments, demonstrating a 71.1% in-training validation success rate on ALFWorld at optimizer step 130. Its primary strength lies in sequential decision-making and task completion in simulated environments.
Loading preview...
Overview
This model, chanyoungkim/grpo-qwen2.5-1.5b-alfworld-step130, is a 1.54 billion parameter Qwen2.5-Instruct backbone that has undergone Reinforcement Learning (RL) fine-tuning using the GRPO (Generative Reinforcement Learning with Policy Optimization) method. The training was conducted on the ALFWorld environment, a text-based interactive simulation designed for evaluating embodied agents.
Key Capabilities
- Agentic Task Completion: Specifically fine-tuned for sequential decision-making and task execution within the ALFWorld environment.
- RL Optimization: Utilizes GRPO with specific hyperparameters, including KL loss and invalid action penalties, to enhance performance in interactive tasks.
- Performance at Step 130: Achieved an in-training validation success rate of 71.1% on a 32-task validation batch at optimizer step 130.
- Resumable Training: Includes
training_state/files, allowing for the resumption of RL training from this exact optimizer step, provided aworld_size=2setup.
Good For
- Research in RL for LLMs: Ideal for researchers exploring GRPO and similar RL fine-tuning techniques for language models in interactive environments.
- ALFWorld Benchmarking: Suitable for evaluating agent performance on ALFWorld tasks, particularly for comparing against other RL-tuned models.
- Developing Agentic Systems: Can serve as a foundation for building and experimenting with language-based agents in simulated environments.