devrim/goodwiki_long_synthetic_ir

GoodWiki-Long-Synthetic (IR format)

A long-document, document-to-document retrieval dataset for training and evaluating retrievers over multi-thousand-word inputs, published in the canonical BEIR/MTEB tri-config layout (queries, corpus, default qrels).

Queries are complete English Wikipedia articles taken from GoodWiki and filtered to long-form content: every query exceeds 16,000 characters, and the pool averages 5,123 words / 6,912 tokens per document (median 4,305 words / 5,817 tokens, tiktoken cl100k_base). The corpus holds model-generated documents about those articles: one fully relevant rephrased counterpart per query, plus two partially relevant topical distractors that centre a different article and weave the query's subject through it. Relevance is therefore graded (score 2 / 1 / 0), which makes the dataset usable both as contrastive training data with hard negatives and as a ranking-robustness benchmark.

The asymmetry is deliberate: the long-context burden sits on the query side (thousands of tokens), while corpus documents are shorter (median 1,726 tokens), which is the regime that document-level retrieval systems have to handle.

Task	Document-to-document retrieval (graded relevance)
Language	English
Queries	17,854 long-form Wikipedia articles
Corpus	53,562 model-generated documents (17,854 fully relevant + 35,708 distractors)
Qrels	53,562 judgements across three query-disjoint splits
Relevance	2 = fully relevant rephrasal · 1 = topical distractor · 0 = everything else
Query length	mean 6,912 / median 5,817 / p95 13,840 / max 35,927 tokens
Corpus doc length	mean 1,815 / median 1,726 / p95 2,527 / max 32,364 tokens

Length statistics are measured on the released files with tiktoken cl100k_base, computed on the decoded document text (see Usage). The paper's Table 1 reports slightly different token statistics for the corpus side, consistent with measurement over the raw serialized strings rather than the decoded text.

Dataset structure

Three configs. queries and corpus are single shared pools with one split each; splits exist only on the qrels config, so a query from any split is scored against the entire corpus and documents belonging to other queries act as hard negatives.

config	split	rows
queries	queries	17,854
corpus	corpus	53,562
default	train	30,000
default	val	6,000
default	test	17,562

queries

field	type	description
_id	string	Wikipedia article id; referenced by query-id in qrels and by reference_article_id in the corpus
text	string	Full article body in structured markdown, opening with a # title heading
title	string	Article title
description	string	One-line article description
categories	list[string]	Wikipedia categories of the article

corpus

field	type	description
_id	string	Document id of the form req-<hash>-<i>-<k>; k=1 is the fully relevant document, k=2 and k=3 its two distractors
text	string	Model-generated document, stored as a JSON-encoded string literal (decode with json.loads, see Usage)
article_type	string	pair (17,854 rows, fully relevant) or distractor (35,708 rows, partially relevant)
reference_article_id	string	_id of the query article this document is judged against
other_article_id	string	_id of the second article involved in generation (always another article from the query pool; 16,957 distinct values)
primary_topic	string	Title + description of the document's main subject: the reference article for pair rows, the other article for distractor rows
integration_topic	string	Title + description of the secondary subject woven into the document: the other article for pair rows, the reference article for distractor rows

Every query article has exactly three corpus documents — one pair and two distractor rows — and every corpus document appears in exactly one qrels row.

default (qrels)

field	type	description
query-id	string	Reference to queries._id
corpus-id	string	Reference to corpus._id
score	int64	2 if the document is the fully relevant rephrasal, 1 if it is a topical distractor

split	unique queries	qrels rows	score=2	score=1
train	10,000	30,000	10,000	20,000
val	2,000	6,000	2,000	4,000
test	5,854	17,562	5,854	11,708

The three splits are query-disjoint and together cover all 17,854 queries. Each query carries exactly three judgements: one at score=2 and two at score=1. A query's split membership is defined solely by which qrels split it appears in.

Usage

from datasets import load_dataset

queries = load_dataset("devrim/goodwiki_long_synthetic_ir", "queries", split="queries")
corpus  = load_dataset("devrim/goodwiki_long_synthetic_ir", "corpus",  split="corpus")

train = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="train")
val   = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="val")
test  = load_dataset("devrim/goodwiki_long_synthetic_ir", "default", split="test")

Corpus documents are stored as JSON string literals (surrounding quotes, escaped newlines and unicode). Decode them before use — this holds for all 53,562 rows:

import json

corpus = corpus.map(lambda r: {"text": json.loads(r["text"])})

Assembling BEIR-style qrels and pools for evaluation:

from collections import defaultdict

qrels = defaultdict(dict)
for row in test:
    qrels[row["query-id"]][row["corpus-id"]] = row["score"]

query_pool = {r["_id"]: r["text"] for r in queries if r["_id"] in qrels}
doc_pool   = {r["_id"]: json.loads(r["text"]) for r in corpus}   # full pool, all splits

Relevance semantics and intended use

Labels are graded rather than binary:

• score = 2 — fully relevant. A rephrased counterpart of the query article, covering the same subject. This is the single correct answer for the query.
• score = 1 — partially relevant. A topical distractor: a document primarily about a different article, with the query's subject integrated into it. It shares real content with the query but is not the right answer.
• score = 0 — irrelevant. Every other document in the pool (the default; these are not materialised as rows).

Training. The fully relevant document is the positive of a contrastive pair; the two distractors enter as partial matches at a reduced target weight rather than as hard negatives with a fully negative target, and the remaining in-batch documents are negatives. This gives every anchor a graded three-way supervision signal — full match, partial match, negative — instead of the usual positive/negative split.

Evaluation. Graded metrics (nDCG with the standard 2^rel - 1 gain) let distractors earn partial credit, so a run is measured not only on whether it retrieves the correct document, but on whether it ranks near-miss documents above unrelated ones. Because the corpus is a single shared pool, every query competes against all 53,562 documents, including those attached to queries in other splits.

Construction

1. Source selection. Long-form English Wikipedia articles from GoodWiki, a cleaned Wikipedia release in structured markdown, filtered to articles above 16,000 characters (roughly a four-thousand-token floor). These become the query pool unchanged, with their original text, title, description and categories.
2. Fully relevant document generation. For each query article, GPT-4o-mini writes a rephrased document whose main subject is that article, with a second, unrelated article named as a secondary theme that the text typically weaves in as framing or analogy. The result remains about the query's subject and is labelled score = 2.
3. Topical distractor generation. For each query article, two further documents are generated with the roles reversed: another article from the pool is the main subject and the query's article is the integrated secondary theme. These share genuine content with the query without answering it, and are labelled score = 1. The shared generation recipe is what makes the distractors topically entangled with the query rather than randomly sampled.
4. Validation. The assembled dataset is checked for duplicate ids in either pool and for referential integrity of every qrels row against both pools; the released files pass with unique ids throughout and no dangling references.
5. IR assembly. Documents are collected into one shared corpus, source articles into one shared query pool, and judgements into query-disjoint train / val / test qrels — the tri-config layout above.

All corpus text is model output, not human-written Wikipedia prose. Only the queries config contains original article text.

Provenance and licensing

Query text comes from English Wikipedia by way of GoodWiki (source), a cleaned markdown release of Good and Featured English Wikipedia articles. Wikipedia article text is licensed CC BY-SA 4.0, and this dataset is released under the same license so that share-alike and attribution are preserved downstream. The GoodWiki extraction tooling and its own packaging are separately MIT-licensed by their author.

Corpus documents are machine-generated with GPT-4o-mini and are marked as synthetic through the article_type field and this card; they are model output about Wikipedia subjects, not verbatim Wikipedia content, and they are not a factual reference. Beyond what is already public in Wikipedia articles, the dataset contains no personal or personally identifying information, and no human subjects were involved in its creation.

Limitations and considerations

• Synthetic corpus text. Rephrased and distractor documents carry the stylistic and factual artifacts of the generating model — reformulated emphasis, occasional embellishment, and drift from the source article's facts. They should not be treated as an accurate account of their subject.
• Not a clean paraphrase. A score = 2 document is a condensed restatement, not a length-matched paraphrase: it is far shorter than its source article (median 1,726 vs. 5,817 tokens) and it carries a woven-in secondary theme from an unrelated article, as recorded in integration_topic.
• Heuristic partial labels. The score = 1 label follows from how a document was constructed (the query's article was the injected secondary theme), not from human judgement of topical closeness. The actual degree of overlap varies, and unlabelled documents elsewhere in the pool may be more related to a query than its own distractors are.
• Lexical leakage. Because a fully relevant document restates its source article, it retains considerable surface overlap with the query, which flatters lexical matching and inflates absolute scores relative to harder retrieval benchmarks.
• Text encoding. Corpus text values are JSON string literals rather than plain strings and must be decoded; skipping this leaves escape sequences and enclosing quotes in the input.
• Coverage. English only, and Wikipedia's Good and Featured articles skew toward well-documented, encyclopedically popular subjects.

Citation

@inproceedings{cavusoglu2026reign,
  title     = {{REIGN}: Refurbished Embeddings with Integrated Guidance Networks for Efficient Context-Length Scaling},
  author    = {{\c{C}}avu{\c{s}}o{\u{g}}lu, Devrim and Akba{\c{s}}, Emre},
  booktitle = {Findings of the Association for Computational Linguistics: {EMNLP} 2026},
  year      = {2026},
  publisher = {Association for Computational Linguistics},
  note      = {To appear}
}

Links

• Code: https://github.com/devrimcavusoglu/reign
• Project page: https://devrimcavusoglu.github.io/reign
• Paper: arXiv link coming soon