snap-stanford/humanual-book

---
license: cc-by-nc-4.0
language:
- en
tags:
- user-simulation
- humanlm
- persona
- e-commerce-and-product-reviews
pretty_name: Humanual-Book
size_categories:
- 10K<n<100K
---

# Humanual-Book

[](https://humanlm.stanford.edu)
[](https://humanlm.stanford.edu/HumanLM_paper.pdf)
[](https://github.com/zou-group/humanlm)
[](https://huggingface.co/collections/snap-stanford/humanual-datasets)

Amazon book reviews from frequent customers, expressing satisfaction or dissatisfaction with book content and reflecting users' preferences and tastes. This dataset is part of the **[HumanLM](https://humanlm.stanford.edu)** benchmark for training user simulators that accurately reflect real user behavior.

**Source:** Amazon Reviews 2023 · **Domain:** E-commerce & Product Reviews · **Date Range:** 1998-01-25 to 2023-05-10

The dataset contains **36,622** comments from **209** users across **33,649** posts, with an average of **1.00** turns per conversation. Each example includes the user's persona, conversation context, and ground-truth response.

**Splits:** train (34,170) · val (492) · test (1,960)

| Column | Description |
|--------|-------------|
| `prompt` | Product description and context as a list of messages with `role` and `content` fields |
| `completion` | The ground-truth user review to generate |
| `persona` | User's review history and preferences (past ratings, writing style) |
| `post_id` | Amazon product ASIN |
| `user_id` | Hashed Amazon reviewer ID (for privacy) |
| `timestamp` | Unix timestamp (milliseconds) of when the review was posted |
| `turn_id` | Always 1 (single-turn reviews) |
| `metadata` | Review metadata as JSON (rating, title, verified purchase, etc.) |

## Quick Start

```python
from datasets import load_dataset

dataset = load_dataset("snap-stanford/humanual-book")
sample = dataset["train"][0]

print(sample["persona"]) # User persona
print(sample["prompt"]) # Conversation context
print(sample["completion"]) # Ground-truth response
```

## Citation

```bibtex
@article{wu2026humanlm,
title={HUMANLM: Simulating Users with State Alignment Beats Response Imitation},
url={https://humanlm.stanford.edu/},
author={Wu, Shirley and Choi, Evelyn and Khatua, Arpandeep and Wang, Zhanghan and He-Yueya, Joy and Weerasooriya, Tharindu Cyril and Wei, Wei and Yang, Diyi and Leskovec, Jure and Zou, James},
year={2026}
}
```

Released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).