xiamoent/Agent-G2-alfworld-1.5b
xiamoent/Agent-G2-alfworld-1.5b is a 1.5 billion parameter language-agent checkpoint, initialized from Qwen2.5-1.5B-Instruct and post-trained with Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning. This model is specifically optimized for the ALFWorld text environment, demonstrating high success rates across various task groups. It features a 32,768-token context length and is designed for research in long-horizon language agents and agentic reinforcement learning within sandboxed environments.
Loading preview...
Agent-G2 ALFWorld 1.5B Overview
xiamoent/Agent-G2-alfworld-1.5b is a specialized language-agent model, built upon the Qwen2.5-1.5B-Instruct architecture and enhanced through Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning. This 1.5 billion parameter model is uniquely designed for the ALFWorld text environment, utilizing an adaptive Gaussian distribution to sample expert-prefix depth for tasks, improving policy optimization without additional probe rollouts.
Key Capabilities
- ALFWorld Specialization: Achieves high success rates in the ALFWorld text environment, including 96.8% for 'Pick' tasks and 95.3% across all tasks.
- Agentic Reinforcement Learning: Incorporates Gaussian Guidance for Agentic Reinforcement Learning (GRPO) during post-training.
- Context Length: Supports a configured context length of 32,768 tokens, with training recipes using up to 4,096 prompt tokens and 512 response tokens.
- Structured Output: Generates responses in a specific
<think>...</think><action>...</action>format, crucial for interaction with the ALFWorld environment.
Good For
- Reproducing Agent-G2 research results in the ALFWorld text environment.
- Research into long-horizon language agents and agentic reinforcement learning.
- Studying adaptive expert-prefix guidance mechanisms.
- Evaluating action selection within environment-provided admissible action sets.