adaamko's picture
Add the GroundLM 2026 (EMNLP) citation
6da23fb verified
|
Raw History Blame Contribute Delete
6.87 kB
---
license: cc-by-4.0
task_categories:
- token-classification
- text-classification
language:
- en
- de
- fr
- es
- it
- ca
- cs
- fi
- sv
- ar
- fa
- hi
- zh
- eu
tags:
- hallucination-detection
- rag
- multilingual
- lettucedetect
size_categories:
- 10K<n<100K
---
# LettuceDetect Prose Hallucination Dataset
Token-level hallucination annotations on LLM answers grounded in **prose**
context, drawn from two public RAG hallucination resources and mapped into one
**unified taxonomy**. This is the prose counterpart to the structured-context
([code, tool output, documents](https://huggingface.co/datasets/KRLabsOrg/lettucedetect-code-hallucination))
collection — together they let a single detector be trained across modalities.
Two sources sit side by side, distinguished by the `dataset` field:
| `dataset` | Spans | Source |
|---|---|---|
| `psiloqa` | **natural** — produced by real LLMs answering Wikipedia-grounded questions | [s-nlp/PsiloQA](https://huggingface.co/datasets/s-nlp/PsiloQA) |
| `ragtruth` | annotated on LLM RAG answers (QA, summarization, data-to-text) | [RAGTruth](https://github.com/ParticleMedia/RAGTruth) |
PsiloQA matters because its hallucinations are **not synthetically injected** —
they occur naturally when models answer grounded questions. That makes this
collection a check that a detector generalizes beyond any single corruption
process, and it spans **14 languages**.
## Statistics
| Split | Total | Hallucinated | Clean |
|-------|-------|-------------|-------|
| train | 78,882 | 63,349 | 15,533 |
| validation | 3,355 | 2,993 | 362 |
| test | 5,597 | 3,563 | 2,034 |
| **Total** | **87,834** | **69,905** | **17,929** |
Per source:
| Source | Total | Languages |
|---|---|---|
| `psiloqa` | 70,044 | 14 |
| `ragtruth` | 17,790 | en |
RAGTruth has no native validation split, so `validation` is PsiloQA only.
## Hallucination Taxonomy
Spans are labeled with the **LettuceDetect unified taxonomy** — three injectable
top-level categories, each with subtypes. Both sources map into the same space.
| Category | Definition | Example subtypes |
|---|---|---|
| `contradiction` | Conflicts with what the context states (wrong value, number, date, entity, relation) | `numerical`, `temporal`, `entity`, `relational`, `value` |
| `unsupported_addition` | Adds a claim the context neither states nor implies | `claim`, `elaboration`, `subjective` |
| `fabricated_reference` | References a named element absent from the context | `entity`, `section`, `identifier` |
**How each source is typed:**
- **RAGTruth** maps mechanically from its native labels (Evident/Subtle Conflict
→ `contradiction`; Baseless Info → `unsupported_addition`, or
`fabricated_reference` when the span introduces a proper noun absent from the
context).
- **PsiloQA** ships only binary character spans with no error type, so each span
is classified into the taxonomy by an LLM that reads the context and the
answer with the span marked. Clean (non-hallucinated) PsiloQA answers are kept
as natural negatives.
## Data Format
Each line in the JSONL files is one sample:
```json
{
"prompt": "User request: ...\n\n...grounded passage...",
"context": "...grounded passage...",
"question": "the user request",
"answer": "the LLM answer (hallucinated or clean)",
"labels": [
{"start": 23, "end": 41, "label": "contradiction",
"category": "contradiction", "subcategory": "numerical"}
],
"split": "train",
"task_type": "qa",
"dataset": "psiloqa",
"language": "en",
"context_modality": "prose",
"category": "contradiction",
"subcategory": "numerical",
"metadata": "{\"id\": \"...\", \"is_hallucinated\": true, \"classifier_model\": \"Qwen/Qwen3.6-35B-A3B\"}"
}
```
**Fields:**
- `prompt` — the full model input, request **first**
(`User request: {question}\n\n{context}`) so it survives context truncation
- `context` / `question` — the grounding passage and the request, separately
(`question` is `null` for RAGTruth summarization). PsiloQA's foreign-language
questions are kept verbatim.
- `answer` — the LLM response (hallucinated or clean)
- `labels` — annotated hallucination spans (empty for clean samples)
- `category` / `subcategory` — unified taxonomy fields (sample-level is a
majority vote over its spans)
- `dataset` — source discriminator (`psiloqa` or `ragtruth`)
- `metadata` — a **JSON-encoded string** of source-specific provenance. Parse
with `json.loads`.
## Usage
```python
import json
from datasets import load_dataset
ds = load_dataset("KRLabsOrg/lettucedetect-prose-hallucination")
# Everything, or filter by source / language
psiloqa = ds["train"].filter(lambda r: r["dataset"] == "psiloqa")
german = ds["train"].filter(lambda r: r["language"] == "de")
for sample in ds["train"]:
if sample["labels"]:
meta = json.loads(sample["metadata"])
print(sample["language"], sample["category"], meta.get("is_hallucinated"))
```
## Limitations
PsiloQA span types are assigned by an LLM rather than human annotators, so the
`(category, subcategory)` labels carry model noise; the underlying spans are from
the original human/automatic PsiloQA annotation. RAGTruth subcategories are
derived from coarse native labels and are not exhaustive.
## Citation
This dataset is the prose part of the unified benchmark in:
```bibtex
@inproceedings{kovacs-etal-2026-beyond,
title = "Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents",
author = "Kov{\'a}cs, {\'A}d{\'a}m and
He, Bowei and
Liu, Xue and
Boros, Istv{\'a}n and
T{\'o}th, Szilveszter and
Recski, Gabor",
editor = "Wang, Yimu and
Choi, Yee Man and
Wu, Di and
Ouyang, Siqi and
Azadani, Mozhgan Nasr and
Fung, Yi R. (May)",
booktitle = "Proceedings of the 1st Workshop on Grounding Language Models: Learning Faithfully and Efficiently ({G}round{LM} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.groundlm-1.44/",
pages = "566--577",
ISBN = "979-8-89176-478-1"
}
```
Also on arXiv: [2607.00895](https://arxiv.org/abs/2607.00895). LettuceDetect itself:
```bibtex
@misc{Kovacs:2025,
title={LettuceDetect: A Hallucination Detection Framework for RAG Applications},
author={Ádám Kovács and Gábor Recski},
year={2025},
eprint={2502.17125},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2502.17125},
}
```
Please also cite the original [PsiloQA](https://huggingface.co/datasets/s-nlp/PsiloQA)
and [RAGTruth](https://github.com/ParticleMedia/RAGTruth) sources.
## License
[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)