alexandrainst/ragtruth-translated-hallucinations

RAGTruth Translated Hallucinations

Multilingual machine translation of RAGTruth into 30 European languages, preserving RAGTruth's word-level hallucination-span annotations. RAGTruth is a corpus of LLM responses to retrieval-augmented generation (RAG) tasks in which humans marked the exact spans that are hallucinated (unsupported by, or contradicting, the provided context). Here both the RAG prompt and the response are translated into each target language, and the annotated hallucination spans are carried over into the translated text.

The dataset is built for and used by EuroEval as a multilingual hallucination-detection benchmark. Compact per-language evaluation subsets are published separately as EuroEval/ragtruth-translated-hallucinations-<lang>-mini (1,000 test / 256 validation examples each).

The translation was done with this script.

Languages

30 languages (one configuration each):

Language	Code	Language	Code	Language	Code
Albanian	sq	Belarusian	be	Bosnian	bs
Bulgarian	bg	Catalan	ca	Croatian	hr
Czech	cs	Danish	da	Dutch	nl
Estonian	et	Faroese	fo	Finnish	fi
French	fr	German	de	Greek	el
Hungarian	hu	Icelandic	is	Italian	it
Latvian	lv	Lithuanian	lt	Norwegian	no
Polish	pl	Portuguese	pt	Romanian	ro
Serbian	sr	Slovak	sk	Slovene	sl
Spanish	es	Swedish	sv	Ukrainian	uk

Usage

from datasets import load_dataset

ds = load_dataset("alexandrainst/ragtruth-translated-hallucinations", "da")

Dataset structure

Each configuration has a single train split (~17,700 examples); use the split column to recover the original RAGTruth train/test partition.

Data fields

Field	Type	Description
prompt	string	The translated RAG prompt (task instruction + retrieved context/source).
answer	string	The translated model response. Hallucinated spans are wrapped in literal <HAL>...</HAL> tags.
labels	list of {start, end, label}	One entry per hallucinated span. start/end are character offsets into answer, pointing at the text inside the tags (start is the index just after the opening <HAL>, end the index just before the closing </HAL>). label is the hallucination type. Empty for non-hallucinated responses.
split	string	The original RAGTruth split: train or test.
task_type	string	The RAG task: QA, Summary, or Data2txt.
dataset	string	Source dataset name (always ragtruth).
language	string	ISO code of the target language.

Roughly 43% of responses contain at least one hallucinated span; the rest are faithful responses with an empty labels list.

Hallucination label types

The four RAGTruth annotation types are preserved:

• Evident Conflict – directly contradicts the provided context.
• Subtle Conflict – contradicts the context in a subtle/indirect way.
• Evident Baseless Info – information clearly not grounded in the context.
• Subtle Baseless Info – ungrounded information that is harder to spot.

Example (da)

{
  "answer": "... <HAL>menes at være døde før den 7. februar 2022.</HAL> ...",
  "labels": [{"start": 560, "end": 603, "label": "Evident Baseless Info"}],
  "split": "train",
  "task_type": "Summary",
  "dataset": "ragtruth",
  "language": "da",
}
# answer[560:603] == "menes at være døde før den 7. februar 2022."

Dataset creation

Source examples come from RAGTruth. Before translation, each annotated hallucination span in the response is wrapped in <HAL>...</HAL> tags. The prompt and the tagged response are then translated into the target language with OpenAI gpt-4o-mini, which is instructed to translate the text including the words inside the tags while copying the tags verbatim, so the annotations transfer to the target language. After translation the tags are parsed back into start/end character offsets.

To keep the gold labels reliable, any sample whose translated tags do not align one-to-one with the original annotated spans (e.g. the model invented, dropped, or unbalanced tags) is discarded rather than kept with fabricated or missing spans.

Considerations & limitations

• Machine-translated. Content is translated by gpt-4o-mini and not post-edited by humans; translation errors and minor span-boundary drift are possible.
• Lower-resource languages (e.g. Faroese, Icelandic) tend to be lower quality than high-resource languages under gpt-4o-mini.
• The hallucination annotations themselves are RAGTruth's original human annotations, projected onto the translated text via the tag mechanism above.

Licensing

The dataset is released under permissive, Apache-2.0-style terms, with one restriction that follows from the translated text being generated by OpenAI's gpt-4o-mini:

You may not use this data to develop artificial-intelligence models that compete with OpenAI's products and services — unless those models are primarily intended to categorize, classify, or organize data (e.g. embeddings or classifiers) and are not distributed or made commercially available to third parties.

For reference, the underlying terms are:

• The upstream RAGTruth corpus is released under the MIT License.
• The translated text consists of outputs from OpenAI's gpt-4o-mini and is governed by the OpenAI Services Agreement, which is the source of the competitive-use restriction above.

Citation

If you use this dataset, please cite the original RAGTruth paper:

@inproceedings{niu2024ragtruth,
  title={Ragtruth: A hallucination corpus for developing trustworthy retrieval-augmented language models},
  author={Niu, Cheng and Wu, Yuanhao and Zhu, Juno and Xu, Siliang and Shum, KaShun and Zhong, Randy and Song, Juntong and Zhang, Tong},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={10862--10878},
  year={2024}
}

Source & acknowledgements

• Original corpus: ParticleMedia/RAGTruth (MIT).
• Translations produced with gpt-4o-mini.
• Translated and uploaded by Freja Thoresen (freja.thoresen@alexandra.dk) and Dan Saattrup Smart (dan.smart@alexandra.dk) at the Alexandra Institute, for EuroEval.
• This research was funded by the EU Horizon project TrustLLM (grant agreement number 101135671).