Publish train-only EnokiQA dataset

README.md

@@ -0,0 +1,162 @@
+---
+pretty_name: EnokiQA
+language:
+  - en
+license: cc-by-sa-4.0
+task_categories:
+  - question-answering
+  - text-generation
+tags:
+  - enoki
+  - hallucination-detection
+  - factuality
+  - long-form-question-answering
+  - wikipedia
+  - parametric-knowledge
+  - datasets
+size_categories:
+  - 10K<n<100K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: data/train-*.parquet
+---
+
+# EnokiQA
+
+![Enoki banner](./enoki-banner.png)
+
+EnokiQA is a dataset of long-form factual questions, no-context answers
+generated from the parametric knowledge of seven language models, and the
+Wikipedia evidence associated with each question.
+
+This release contains the **19,594-example unannotated train split** in its
+natural generator and Wikipedia-popularity distribution. It is intended for
+research on long-form question answering, factuality, hallucination detection,
+and evidence-based verification.
+
+- Paper: [Enoki: Efficient Multi-Level Hallucination Detection](https://arxiv.org/abs/2609.00581)
+- Code: [s-nlp/Enoki](https://github.com/s-nlp/Enoki)
+- Original release: [Harvard Dataverse](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/6TN4ZM)
+
+## Loading the dataset
+
+Install 🤗 Datasets:
+
+```bash
+pip install datasets
+```
+
+Load the complete train split:
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("s-nlp/EnokiQA", split="train")
+print(dataset)
+print(dataset[0])
+```
+
+Stream examples without downloading the full Parquet file:
+
+```python
+dataset = load_dataset("s-nlp/EnokiQA", split="train", streaming=True)
+
+for example in dataset.take(3):
+    print(example["question"])
+```
+
+Filter by answer generator:
+
+```python
+qwen3_8b = dataset.filter(
+    lambda example: example["answer_model"] == "Qwen_Qwen3-8B"
+)
+```
+
+## Dataset structure
+
+Each row contains a question, a generated answer, two levels of Wikipedia
+evidence, generator metadata, and article-popularity statistics.
+
+| Field | Type | Description |
+|---|---|---|
+| `id` | string | Stable example identifier |
+| `split` | string | Original split; always `train` in this release |
+| `title` | string | Wikipedia article title |
+| `question` | string | Long-form factual question |
+| `answer` | string | No-context model-generated answer |
+| `answer_model` | string | Generator model identifier |
+| `answer_length` | int32 | Answer length in characters |
+| `paragraph_context` | string | Wikipedia paragraph used to generate the question |
+| `full_page_context` | string | Full Wikipedia article text |
+| `context_id` | string | Stable paragraph-context hash |
+| `wiki_url` | string | Wikipedia article URL |
+| `wiki_pageid` | int64 | Wikipedia page ID |
+| `wiki_categories` | list[string] | Wikipedia article categories |
+| `wiki_qid` | string | Wikidata QID |
+| `wiki_article_length` | int32 | Source-reported Wikipedia article length |
+| `pv_mean` | float64 | Mean daily pageviews over the collection window |
+| `pv_total` | int64 | Total pageviews over the collection window |
+| `pv_p50` | int64 | Median daily pageviews |
+| `pv_p95` | int64 | 95th percentile daily pageviews |
+| `popularity_tier` | string | `low`, `medium`, or `high` popularity bucket |
+
+## Generator models
+
+The answers were produced by seven instruction-tuned language models:
+
+- Qwen2.5 7B, 14B, and 32B Instruct
+- Qwen3 4B and 8B
+- Llama 3.1 8B Instruct
+- Mixtral 8x7B Instruct
+
+## Dataset statistics
+
+| Statistic | Train |
+|---|---:|
+| Examples | 19,594 |
+| Unique Wikipedia contexts | 2,226 |
+| Generator models | 7 |
+| Average answer length | 5,525 characters |
+| Median answer length | 4,773 characters |
+| Average full-page context | 13,034 characters |
+
+### Wikipedia popularity
+
+| Tier | Examples |
+|---|---:|
+| Low, fewer than 100 views/day | 11,423 |
+| Medium, 100–1,000 views/day | 7,079 |
+| High, more than 1,000 views/day | 1,092 |
+
+## Intended use and limitations
+
+- This train-only release does not contain hallucination labels.
+- Answers were generated without retrieval context; they reflect the
+  generators' parametric knowledge and their model-specific biases.
+- Wikipedia evidence may be incomplete or temporally outdated.
+- Generator representation is not balanced in the natural train split.
+- `full_page_context` can be long; use streaming or column selection when only
+  question-answer pairs are needed.
+
+## License
+
+Wikipedia-derived content and the accompanying dataset layer are distributed
+under CC BY-SA 4.0. Users should preserve attribution and share-alike
+requirements when redistributing derived datasets.
+
+## Citation
+
+```bibtex
+@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},
+}
+```

data/train-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e34d5539971bc74ce36fc990ddb971426f8287efa6c21a5c484e9235ebef82cd
+size 72566467