facebook/gistbench

---
license: cc-by-nc-4.0
task_categories:
- text-classification
language:
- en
size_categories:
- 1M<n<10M
---

# GISTBench — Groundedness & Interest Specificity Test Bench

GISTBench evaluates how well LLMs understand users from their engagement history. Given a user's interactions with content (videos, articles, books, etc.), the benchmark measures whether an LLM can extract meaningful interests, ground them in evidence, and cite specific relevant items.

## Dataset Details

- **Rows:** 4,214,059 engagements
- **Users:** ~1,000 anonymized users
- **Format:** Parquet

### Schema

| Column | Type | Description |
|---|---|---|
| `user_id` | `int64` | Anonymized user identifier |
| `object_id` | `int64` | Anonymized content item identifier |
| `object_text` | `string` | Text description of the content item |
| `interaction_type` | `string` | One of: `explicit_positive`, `implicit_positive`, `implicit_negative`, `explicit_negative` |
| `interaction_time` | `string` | Anonymized interaction timestamp |

### Interaction Types

| Type | Meaning | Examples |
|---|---|---|
| `explicit_positive` | User actively expressed positive interest | Liked, favorited, rated highly |
| `implicit_positive` | Passive positive signal | Watched fully, clicked through |
| `implicit_negative` | Passive negative signal | Scrolled past, skipped |
| `explicit_negative` | User actively expressed dislike | Downvoted, reported, rated low |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("facebook/gistbench")
print(ds)
# DatasetDict({
# train: Dataset({
# features: ['interaction_type', 'user_id', 'object_id', 'interaction_time', 'object_text'],
# num_rows: 4214059
# })
# })

# Access rows
print(ds["train"][0])
```

## Citation

If you use GISTBench in your research, please cite:

```bibtex
@misc{fostiropoulos2026gistbench,
title={GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification},
author={Iordanis Fostiropoulos and Muhammad Rafay Azhar and Abdalaziz Sawwan and Boyu Fang and Yuchen Liu and Jiayi Liu and Hanchao Yu and Qi Guo and Jianyu Wang and Fei Liu and Xiangjun Fan},
year={2026},
eprint={2603.29112},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2603.29112}
}
```

## License

This dataset is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).