---
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - webshop
  - world-model
  - w2r
  - tm-wm
  - qwen3-8b
  - step92
  - context-40k
pretty_name: WebShop TM-WM Step92 WM/W2R Trajectories - Qwen/Qwen3-8B (40k, TP=1)
---

# WebShop TM-WM Step92 WM/W2R Trajectories - Qwen/Qwen3-8B (40k context)

Trajectories for the WebShop world model `Ricardo-H/ws-llama-webshop-token-match-0429-step-92`, evaluated with `Qwen/Qwen3-8B` as the agent. Both WM and agent use vLLM with `max_model_len=40960`, eliminating the agent context overflow seen at 32k.

## Configuration

- World model: `Ricardo-H/ws-llama-webshop-token-match-0429-step-92`
- World model serving: vLLM, tensor parallel size 2, `max_model_len=40960`
- Agent: `Qwen/Qwen3-8B`
- Agent serving: vLLM, tensor parallel size 1, `max_model_len=40960`, temperature 0, thinking disabled
- Eval concurrency: 32
- Task: WebShop test split (200 items)

## Metrics

- WM total items: 200, errors: 14, accuracy: 13.98%
- W2R total items: 200, errors: 14, accuracy: 10.22%

## Files

- `webshop_*.json`: WM interaction trajectories.
- `_metrics.json`: WM metrics.
- `valid_on_real_env/webshop_*.json`: W2R replay results.
- `valid_on_real_env/_metrics.json`: W2R metrics.
- `metadata.json`: configuration.