Ricardo-H/

ws-w2w-llama-3.1-8b-webshop-wm-w2r-qwen3-32b

Trajectories for the WebShop SFT baseline world model X1AOX1A/WorldModel-Webshop-Llama3.1-8B (W2W: world-model-only training), evaluated with Qwen/Qwen3-32B as the agent.

World model: X1AOX1A/WorldModel-Webshop-Llama3.1-8B World model serving: vLLM, tensor parallel size 1, max_model_len=32768 Agent: Qwen/Qwen3-32B Agent serving: vLLM, tensor parallel size 4, max_model_len=40960, temperature… See the full description on the dataset page: https://huggingface.co/datasets/Ricardo-H/ws-w2w-llama-3.1-8b-webshop-wm-w2r-qwen3-32b.