xiamoent/Agent-G2-webshop-1.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 2, 2026Architecture:Transformer Featherless Exclusive Cold

xiamoent/Agent-G2-webshop-1.5b is a 1.5 billion parameter language-agent checkpoint developed by xiamoent, specialized for the WebShop simulator. Initialized from Qwen2.5-1.5B-Instruct, it is post-trained with Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning, which adaptively samples expert-prefix depth. This model excels at navigating the WebShop environment, achieving a 92.3 reward score and 78.9% final-purchase success, making it ideal for research in long-horizon language agents and agentic reinforcement learning.

Loading preview...

Overview

xiamoent/Agent-G2-webshop-1.5b is a 1.5 billion parameter language-agent model, built upon Qwen2.5-1.5B-Instruct, specifically designed for the WebShop simulator. It leverages Agent-G2: Gaussian Guidance for Agentic Reinforcement Learning, a post-training method that uses an adaptive Gaussian distribution to sample expert-prefix depth for each task, enhancing policy optimization without additional probe rollouts.

Key Capabilities

  • WebShop Specialization: Optimized for interaction within the sandboxed WebShop e-commerce environment.
  • Agentic Reinforcement Learning: Incorporates Gaussian Guidance for improved agentic behavior and long-horizon task completion.
  • High Performance: Achieves a 92.3 Reward Score and 78.9% Final-purchase Success on the WebShop benchmark.
  • Structured Action Generation: Generates actions in a specific <think>...</think><action>...</action> format, requiring environment-side parsing and validation.
  • Context Length: Configured with a 32,768 token context length, though training used up to 4,096 prompt tokens and 512 response tokens.

Intended Use Cases

  • Reproducing Agent-G2 Results: Ideal for replicating the research findings of the Agent-G2 project within the WebShop simulator.
  • Research on Language Agents: Suitable for studies on long-horizon language agents and agentic reinforcement learning.
  • Adaptive Expert-Prefix Guidance: Useful for investigating the effectiveness and mechanisms of adaptive expert-prefix guidance in agent training.
  • Action Selection Evaluation: Designed for evaluating action selection over environment-provided admissible action sets.

Finetunes
1 models