s-nlp/EnokiQA

EnokiQA

EnokiQA is an annotated dataset for fine-grained hallucination detection in long-form question answering. Each example contains a factual question, a no-context LLM answer, the full Wikipedia article used as verification evidence, sentence-grouped factual triples, and per-triple NLI and hallucination probabilities.

The dataset is dual-granularity: every hallucination label is attached to a claim (an extracted triple) and projected to a character span of the answer. The same annotation therefore supports claim-level verification, sentence-level classification, and span-level localization, without separate annotation layers.

Available splits: dev and test.

• Code: s-nlp/Enoki
• Paper: Enoki: Efficient Multi-Level Hallucination Detection

Loading the dataset

pip install datasets

from datasets import load_dataset

dataset = load_dataset("s-nlp/EnokiQA")
dev = dataset["dev"]
test = dataset["test"]

Inspect hallucinated triples using the canonical hall_prob > 0.5 threshold:

example = test[0]

for sentence in example["sentences"]:
    hallucinated = [
        triple
        for triple in sentence["triples"]
        if triple["hall_prob"] > 0.5
    ]
    if hallucinated:
        print(sentence["sentence"])
        for triple in hallucinated:
            print(triple["triplet"], triple["hall_prob"])

Stream a split without downloading it in full:

test = load_dataset("s-nlp/EnokiQA", split="test", streaming=True)

Dataset structure

Field	Type	Description
id	string	Source example identifier
title	string	Wikipedia article title
question	string	Long-form factual question
answer	string	No-context model-generated answer
context	string	Wikipedia evidence used to verify the answer
n_sentences	int32	Number of annotated answer sentences
n_triplets	int32	Total extracted triples across the answer
sentences	list[Sentence]	Sentence-grouped triple annotations

Sentence

Field	Type	Description
sentence_index	int32	Zero-based position in the answer
sentence	string	Sentence text (matches answer up to whitespace normalization)
hall_prob	float64	Sentence hallucination probability: max(hall_prob) over its triples (0.0 if the sentence has no triples)
triples	list[Triple]	Extracted and verified factual triples

A sentence is classified as hallucinated when its hall_prob > 0.5, i.e. when at least one of its triples is hallucinated.

Triple

A claim in EnokiQA is a triple: each extracted (subject, predicate, object) is one atomic factual statement of the answer.

Field	Type	Description
triplet	list[string] of length 3	[subject, predicate, object]
span	list[int32]	Object's [start, end) character span in answer; empty if the object could not be located
hypothesis	string	Verbalized triple supplied to the NLI verifier
entailment	float64	Probability that the context entails the hypothesis
neutral	float64	Probability that the context is neutral
contradiction	float64	Probability that the context contradicts the hypothesis
hall_prob	float64	Hallucination probability: neutral + contradiction

A triple is classified as hallucinated when hall_prob > 0.5.

Triples within a sentence are incremental: the extractor emits a chain of increasingly specific facts (Dutse | is | city → Dutse | is | capital city → Dutse | is | capital city of Jigawa State), so the object spans of consecutive triples in a sentence are typically nested. The most specific triple whose hall_prob crosses the threshold localizes the unsupported information. See the Evaluation section for how to turn this into span labels.

Dataset statistics

Statistic	Dev	Test
Examples	1,995	1,995
Generator models	7	7
Sentences	58,454	57,987
Triples	499,923	492,268
Hallucinated triples	78.9%	77.2%
Average answer length (chars)	5,636	5,727
Average context length (chars)	13,942	15,816
Average triples per answer	250.6	246.8
Mean per-document hallucination rate	77.8%	76.0%

Both splits are balanced across the seven generator models (285 answers per model). The dev split is additionally stratified to match the test distribution over generator model and Wikipedia popularity tier.

Data construction

1. Article sampling. English Wikipedia articles were sampled across pageview popularity tiers (low / medium / high) to cover both frequent and long-tail entities.
2. Question generation. Paragraph-level contexts were used to generate long-form factual questions with GPT-OSS-120B, followed by an LLM-based filter for grounding, quality, and answerability.
3. Answer generation. Seven instruction-tuned LLMs answered each question without access to the article, relying on parametric knowledge only: Qwen2.5-7B/14B/32B-Instruct, Qwen3-4B, Qwen3-8B, Llama-3.1-8B-Instruct, and Mixtral-8x7B-Instruct.
4. Filtering. Answer relevance filtering, length filtering, and duplicate removal were applied before forming the splits.
5. Evidence. The full Wikipedia article is retained as context for verification.

The generation pipeline is released in the dataset_generation/ directory of the code repository.

Evaluation

Tasks derived from the annotations

All three tasks are derived from the same hall_prob field, so a single model output can be scored at every granularity.

Labels are heavily skewed towards the hallucinated class (about 78% of triples and 94% of sentences), so a constant positive predictor already reaches a positive-class F1 above 0.96. We therefore report macro-F1 (the unweighted mean of the F1 of both classes) and threshold-free AUROC for the classification tasks, following the sentence-level protocol of the paper.

Task	Input	Gold label	Recommended metrics
Sentence-level detection	context, sentence	sentence hall_prob > 0.5 (any of its triples is hallucinated)	macro-F1, AUROC
Claim-level verification	context, hypothesis	hall_prob > 0.5 per triple	macro-F1, AUROC
Span-level localization	context, answer	span of every hallucinated triple	Span Coverage F1 (micro), optionally IoU / exact-match F1

Span Coverage F1

Gold spans are nested and often coarser than the actual error, so exact-match span F1 and IoU penalize precise localization. Span Coverage F1 instead counts a predicted span as correct if it lies inside a gold span: precision is the share of predicted spans covered by some gold span, recall is the share of gold spans that cover some predicted span, and both are micro-averaged over the split.

We use the reference implementation from mycelium.

Minimal scoring example

from datasets import load_dataset
from sklearn.metrics import f1_score, roc_auc_score
from mycelium import evaluate

THRESHOLD = 0.5
test = load_dataset("s-nlp/EnokiQA", split="test")


def sentence_labels(example):
    """1 if the sentence is hallucinated (its hall_prob is the max over its triples)."""
    return [int(s["hall_prob"] > THRESHOLD) for s in example["sentences"]]


def gold_spans(example):
    """Half-open [start, end) spans of hallucinated triples in `answer`."""
    return [
        t["span"]
        for s in example["sentences"]
        for t in s["triples"]
        if t["hall_prob"] > THRESHOLD and t["span"]
    ]


# Replace these two functions with your system's outputs.
def my_sentence_scores(example):
    """Return one hallucination score in [0, 1] per sentence."""
    ...


def my_spans(example):
    """Return predicted hallucinated spans in `answer` as [start, end)."""
    ...


y_true, y_score, gold, pred = [], [], [], []
for ex in test:
    y_true += sentence_labels(ex)
    y_score += my_sentence_scores(ex)
    gold.append(gold_spans(ex))
    pred.append(my_spans(ex))

y_pred = [int(s > THRESHOLD) for s in y_score]
report = evaluate(
    references=gold,
    predictions=pred,
    metrics=["span_coverage", "iou"],
    average="micro",
)
print("sentence macro-F1 ", f1_score(y_true, y_pred, average="macro"))
print("sentence AUROC    ", roc_auc_score(y_true, y_score))
print("span coverage P/R/F1", report["span_coverage"])
print("character IoU", report["iou"]["score"])

Annotation method

The annotations were produced automatically by a common pipeline for both splits:

1. Answers were split into sentences.
2. Enoki-LLM with GPT-OSS-120B extracted incremental (subject, predicate, object) triples from each sentence.
3. Triple object spans were localized in the generated answer when possible.
4. Each verbalized triple (hypothesis) was checked against the full Wikipedia article with a Qwen3.5-9B NLI-style verifier, producing entailment, neutral, and contradiction probabilities.
5. hall_prob was calculated as neutral + contradiction; a triple is hallucinated when hall_prob > 0.5.

Human validation

To assess annotation quality, 100 randomly sampled test examples were labeled independently by two human annotators.

Measure	Value
Human–human raw agreement, character-level accuracy	0.80
Automatic pipeline vs. humans, sentence-level F1	0.87
Automatic pipeline vs. humans, span-level Coverage F1	0.57

License

Wikipedia-derived content and the annotation layer are distributed under CC BY-SA 4.0. Preserve attribution and share-alike requirements when redistributing derived datasets.

Citation

@misc{rykov2026enokiefficientmultilevelhallucination,
  title         = {Enoki: Efficient Multi-Level Hallucination Detection},
  author        = {Elisei Rykov and Timur Ionov and Nikolay Ivanov and Maksim Savkin and Maksim Makarenko and Alexander Panchenko and Vasily Konovalov and Julia Belikova},
  year          = {2026},
  eprint        = {2609.00581},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.00581},
}