configs:
- config_name: alfworld
data_files:
- split: test
path: data/alfworld/**/*.json
- config_name: webshop
data_files:
- split: test
path: data/webshop/**/*.json
- config_name: blocksworld
data_files:
- split: test
path: data/blocksworld/**/*.json
- config_name: scienceworld
data_files:
- split: test
path: data/scienceworld/**/*.json
- config_name: textworld
data_files:
- split: test
path: data/textworld/**/*.json
default_config_name: alfworld
🌟 RewardPrediction: A Fine-grained Step-wise Reward Prediction Benchmark
🌐 Website | 💻 GitHub | 📄 arXiv
RewardPrediction is a large-scale benchmark designed to evaluate fine-grained, step-wise reward prediction across five diverse text-based environments: AlfWorld, ScienceWorld, TextWorld, WebShop, and BlocksWorld. It comprises a total of 2,454 unique trajectories with dense reward annotations.
To prevent heuristic reward hacking, we structured the benchmark using a paired positive-negative strategy:
- Positive Trajectories: Expert demonstrations augmented with random interaction steps at the boundaries.
- Negative Trajectories: Failure trajectories generated via a random policy.
📥 Load RewardPrediction Benchmark
Our benchmark can be loaded from the 🤗 huggingface repo at YijunShen/RewardPrediction (reconstructing the alfworld/, webshop/ folders, etc.).
from huggingface_hub import snapshot_download
import shutil, os; from pathlib import Path
# [Optional] Your Hugging Face token (e.g., "hf_...") to avoid rate limits
HF_TOKEN = None
# 1. Download the raw files from the repository
snapshot_download(
repo_id="YijunShen/RewardPrediction",
repo_type="dataset",
local_dir="rewardprediction",
token=HF_TOKEN
)
# 2. Unwrap 'data' folder to restore the original environment tree
d = Path("rewardprediction/data")
if d.exists():
[shutil.move(str(i), "rewardprediction") for i in d.iterdir()]
d.rmdir()
print(f"✨ Original structure restored at: {os.path.abspath('rewardprediction')}")
📄 Data Schema
Each row in the dataset represents a complete task trajectory. The data features a nested structure to efficiently store sequential interactions:
- goal description (string): The natural language goal the agent needs to achieve for this specific trajectory.
- trajectory (list): A nested sequence of interaction steps. Each step contains the following fields:
- action (string): The specific action executed by the agent at this time step.
- observation (string): The textual feedback/observation returned by the environment.
- reward (dict): A dictionary containing fine-grained reward labels:
raw(float): The native, sparse environment reward (usually 1.0 for success, 0.0 otherwise).shaped(float): The interpolated, step-wise ground-truth reward.is_expert(boolean): Indicates whether this step is part of an expert demonstration.
✍️ Citation
If you find this dataset helpful for your research, please cite our work:
@misc{shen2026StateFactory,
title={Reward Prediction with Factorized World States},
author={Yijun Shen and Delong Chen and Xianming Hu and Jiaming Mi and Hongbo Zhao and Kai Zhang and Pascale Fung},
year={2026},
eprint={2603.09400},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2603.09400},
}