Trajectories for the WebShop world model Ricardo-H/ws-llama-webshop-token-match-0429-step-60, evaluated with Qwen/Qwen3-8B as the agent. Both the WM and the agent are served by vLLM with max_model_len=40960 (Qwen3 native max), which eliminates the agent-side context overflow errors observed at 32k.
World model: Ricardo-H/ws-llama-webshop-token-match-0429-step-60
World model serving: vLLM… See the full description on the dataset page:
https://huggingface.co/datasets/Ricardo-H/ws-step60-webshop-wm-w2r-qwen3-8b-40k.
