Dataset · Qwen
World modeling benchmark for LLM agents
Qwen/AgentWorldBench
Evaluation of linguistic world models using agent trajectories and ground-truth observations from real environments.
What's inside
The benchmark is compiled from real observations of frontier model trajectories on Tool Decathlon, Terminal-Bench 1.0/2.0, and OSWorld-Verified. Each example is paired with a ground-truth observation from execution in the environment, enabling reference-grounded scoring. Modalities include tabular and text in English, with a size of 1K–10K.
What you can do with it
- Evaluating LLM agents on action consequence prediction
- Testing world modeling within an agentic system
- Comparing models using reference-grounded scoring
- Tasks
- Text generation
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 1.3K
- Likes
- 110
Models trained on it
- unsloth/Qwen-AgentWorld-35B-A3B-GGUFText generation 67.1K
- Qwen/Qwen-AgentWorld-35B-A3BText generation 9.9K
- hotdogs/qwen35b-a3b-fable-sft-abliteratedText generation 2.8K
- DhruvalLabs/Qwen-AgentWorld-35B-A3B-GGUFImage text → text 935
- gaoqianshen/Qwen-AgentWorld-35B-A3B-Q4_K_M-GGUFText generation 729
- unsloth/Qwen-AgentWorld-35B-A3BText generation 371
- cyankiwi/Qwen-AgentWorld-35B-A3B-AWQ-FP8Text generation 137
- cyankiwi/Qwen-AgentWorld-35B-A3B-AWQ-INT4Text generation 133
How to load it
from datasets import load_dataset
ds = load_dataset("Qwen/AgentWorldBench")