SUSTech/BCP-Link-corpus

BrowseComp-Plus-Link (BCP-Link) is a benchmark for evaluating how effectively search agents can use hyperlinks to discover evidence beyond initial search results in an offline, fully reproducible search environment. Built on BrowseComp-Plus, it recovers 63,371 verified links among the fixed corpus of 100,195 offline webpages. These links are inserted directly into the document text and exposed through standardized search and visit tools, enabling controlled analysis of whether search agents can recognize useful links, navigate across documents, gather relevant evidence, and reach the correct answer efficiently.

Dataset Introduction

The current release includes:

• 6.5 GB uncompressed corpus data
• 13.8 MB link metadata
• 100,195 offline webpages
• 63,371 verified in-corpus hyperlinks
• 17,633 documents with at least one incoming or outgoing link, representing 17.60% graph participation
• 4.59 outgoing links per document with outgoing links on average

Corpus Schema (corpus)

corpus is the primary configuration. It provides the complete link-enriched document corpus in eight Parquet shards for search, retrieval, and agent evaluation. Each row represents one document:

Field	Type	Description
docid	string	Stable document identifier inherited from BrowseComp-Plus.
url	string	Canonical document URL.
url_raw	string, optional	Source URL before normalization, when available.
text_raw	string	Baseline text without exposed inline links.
text	string	Link-enriched text, ready for direct use.
links	list	Link metadata with to_url and url_text.

Example record:

{
  "docid": "12345",
  "url": "https://example.org/page",
  "text_raw": "Document text.",
  "text": "Document text with [a page link](https://example.org/other).",
  "links": [{"to_url": "https://example.org/other", "url_text": "a page link"}]
}

Link Table Schema (links)

links is the optional configuration. It reads data/links.jsonl and provides link-level metadata for hyperlink-graph, structural, and retrieval analysis. Each row represents one verified link between two corpus documents:

Field	Type	Description
source_docid	string	Document containing the link.
target_docid	string	Matched target document.
source_url	string	Canonical source URL.
target_url	string	Canonical target URL.
anchor_text	string	Text displayed for the link.

How to Use

from datasets import load_dataset

corpus = load_dataset(
    "SUSTech/BCP-Link-corpus",
    "corpus",
    split="train",
)

links = load_dataset(
    "SUSTech/BCP-Link-corpus",
    "links",
    split="train",
)

Corpus Construction

BCP-Link-corpus uses the fixed text released with BrowseComp-Plus as its reference rather than creating a new web snapshot. During recovery, 97,682 of 100,195 source pages (97.5%) provided sufficient content for link reconstruction; all 100,195 documents remain in the release.

1. Page recovery. Corpus URLs are processed in batches. The live page is tried first; if it is unavailable or incomplete, an archived copy is used instead. Pages that depend on client-side rendering are opened in a browser. Cases that still fail automatic recovery are reviewed manually, including pages with unusual layouts and historical Wikipedia revisions.
2. Main-content selection. Navigation menus, advertisements, sidebars, and other page furniture are removed before links are collected. The recovered main content is compared with the fixed BrowseComp-Plus text. Pages with poor text coverage or unreliable content are discarded; accepted pages then pass through the same cleaning and deduplication steps.
3. Link extraction and matching. Links are extracted only from accepted main content. Target URLs are normalized and looked up in the BrowseComp-Plus URL index. BCP-Link-corpus keeps a link only when its target maps to another document in the corpus. Self-links, fragment-only links, links from non-content regions, and links that cannot be matched reliably are excluded.
4. Text alignment. Each retained anchor is aligned with its occurrence in text_raw. The matched span is written as a Markdown link in text, while the corresponding links entry and links.jsonl record are produced at the same time. The final pass checks document-ID uniqueness, target existence, and agreement between inline links and structured metadata.

The Team

BCP-Link is built and maintained by the group of Assistant Professor Wei Hongxin and Professor Jing Bingyi in SUSTech and CUHK-SZ.

Core maintainers: Yu Zhengye, Hao Sai, Shi Shunuo, and Li Hanyang.

Citation

If you use BCP-Link corpus, please cite:

@misc{bcp-link-corpus-2026,
  title        = {BCP-Link-corpus},
  author       = {Wei, Hongxin and Jing, Bingyi and Yu, Zhengye and Hao, Sai and Shi, Shunuo and Li, Hanyang and Zhang, Songxin and Xie, Zejian},
  year         = {2026},
  version      = {v1.0},
  howpublished = {Hugging Face dataset: SUSTech/BCP-Link-corpus},
  url          = {https://huggingface.co/datasets/SUSTech/BCP-Link-corpus}
}

BCP-Link uses the questions and evidence annotations from BrowseComp-Plus. When reporting BCP-Link results, please also follow its official citation instructions.