wckwan/ALFWorld-Qwen3-8B-GiGPO
The wckwan/ALFWorld-Qwen3-8B-GiGPO model is an 8 billion parameter Qwen3-based policy fine-tuned for text-based embodied household tasks within the ALFWorld environment. It utilizes the GiGPO estimator, which combines episode-level and discounted step-level group advantages, to achieve high success rates in complex interactive tasks. This model is specifically optimized for robust performance in sequential decision-making and planning within simulated environments, demonstrating strong capabilities in tasks like 'pick_and_place' and 'pick_heat_then_place'.
Loading preview...
Overview
This model, ALFWorld-Qwen3-8B-GiGPO, is a specialized policy built upon the Qwen3-8B architecture, fine-tuned for the ALFWorld text-based embodied household task environment. It leverages the GiGPO (episode-level group advantages plus discounted step-level group advantages) estimator during training to enhance its performance in complex, multi-step tasks.
Key Capabilities
- Embodied Task Performance: Achieves high success rates in ALFWorld tasks, including
pick_and_place,pick_heat_then_place,pick_cool_then_place,pick_clean_then_place, andpick_two_obj_and_place. - Advanced Reinforcement Learning: Utilizes the GiGPO estimator for robust policy optimization, demonstrating improved performance over standard outcome-based GRPO in later training stages.
- Context Handling: Trained with a
max_prompt_lengthof 4096 tokens, ensuring comprehensive understanding of task descriptions and environmental states without prompt clipping.
Good For
- Research in Embodied AI: Ideal for researchers exploring advanced reinforcement learning techniques and policy optimization in text-based interactive environments.
- ALFWorld Benchmarking: Provides a strong baseline and comparison point for new methods developed for the ALFWorld suite of tasks.
- Sequential Decision Making: Demonstrates strong capabilities in planning and executing multi-step actions based on textual observations.