HAT: Hallucination Annotation for Translation

🧭 Table of Contents

• Overview
• Usage
• Data Creation Process
• Data Statistics
• Dataset Structure
• Paper Abstract
• Citation
• License

---

📘 Overview

HAT (Hallucination Annotation for Translation) is a large-scale dataset for hallucination detection in machine translation (MT).
It is released as part of our publication at ACL 2026 (paper).

• 350,959 span-level annotated samples
• 38 language pairs
• ~8,000-10,000 samples per language pair, divided into train, dev, and test sets
• Annotations by professional translators under strict quality control

HAT enables research on detecting and mitigating hallucinations in MT systems.

---

🚀 Usage

Load the dataset with the 🤗 datasets library:

from datasets import load_dataset

# All 38 language pairs
ds = load_dataset("apple/hat")
train, validation, test = ds["train"], ds["validation"], ds["test"]

# A single language pair (e.g. English → Japanese)
ds = load_dataset("apple/hat", "en_US-ja_JP")

The available config names are the 38 language pairs (e.g. en_US-ja_JP, fr_FR-en_US); the default config concatenates all pairs. Each config exposes train, validation, and test splits.

---

⚙️ Data Creation Process

The dataset was created through the following steps:

1. Data Collection:
   Crawled ~5M sentences per language from the web, filtered by language ID, length, and deduplication.

2. Translation:
   Translated monolingual sentences into target languages using a strong neural MT model.

3. Sample Selection:
   Selected translations with lower quality (based on quality metrics) to increase the likelihood of hallucinations.

4. Annotation:
   Professional translators labeled hallucinations under rigorous quality control.

5. Post-processing:
   Removed samples with issues in source text to ensure data integrity.

---

📊 Data Statistics

Split	Samples per language pair
Train	~10,000
Dev	~2,000
Test	~3,000

---

📂 Dataset Structure

Datasets are organized in the data/ directory:

data/
├── train/
├── dev/
└── test/

Each split contains subdirectories for each language pair which contains one parquet file.

Split naming: the parquet files live under data/train, data/dev, and data/test. On the Hugging Face Hub the dev files are exposed as the standard validation split, so load_dataset(...) returns train / validation / test.

Schema:

Field	Type	Description
source_locale	string	Locale of the source text
source_text	string	Source sentence
target_locale	string	Locale of the target text
target_text	string	Machine-translated output
split	string	Dataset split (train / dev / test)
label	int64	Binary hallucination label (0: no hallucination, 1: contains hallucination)
score	float64	Proportion (0–1) of hallucinated characters
annotation	string	Raw span-level hallucination annotation

📄 Paper Abstract

Hallucinations in machine translation (MT)—outputs that may be fluent yet unfaithful to the source content—remain a critical obstacle. They hinder the reliable deployment of MT systems in real-world applications. Despite growing attention to this phenomenon, progress has been constrained by the lack of large-scale, high-quality benchmarks dedicated to hallucination detection. We introduce HAT (Hallucination Annotation for Translation), a novel dataset designed to advance research on this problem. HAT comprises 350,959 span-level annotated samples across 38 language pairs, with approximately 8,000–10,000 samples per pair partitioned into training, development, and test sets. Annotations were produced by professional translators under rigorous quality control protocols to ensure reliability. We provide a detailed analysis of hallucination distributions and establish benchmark performance using a diverse set of baselines, including automatic MT evaluation metrics as well as large language models. By providing the first large-scale, systematically annotated resource for hallucination detection in MT, HAT enables the development of more faithful translation models and lays the groundwork for future research on building trustworthy machine translation systems.

📚 Citation

If you use this dataset, please cite:

@inproceedings{chatterjee-etal-2026-hat,
    title = "{HAT}: Hallucination Annotation for Translation",
    author = "Chatterjee, Rajen  and
      Li, Xintong  and
      Charoenpornsawat, Paisarn  and
      Lee, Allen",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.721/",
    pages = "15865--15888",
    ISBN = "979-8-89176-390-6",
}

📄 License

The HAT dataset is licensed under CC BY-NC-ND 4.0 (https://creativecommons.org/licenses/by-nc-nd/4.0/).