mast-benchmark/100k-corpus-2026

MAST 100K Corpus 2026

This dataset contains the fixed English document corpus used for MAST @ FIRE 2026, the Multilingual Agentic Search Track. MAST evaluates whether multilingual agentic search systems can answer complex questions posed in different languages by retrieving English evidence and producing short, correct English answers.

This corpus is copied from BrowseComp-Plus, a benchmark for Deep-Research systems that isolates the effect of the retriever and the LLM agent to enable fair and transparent comparisons. BrowseComp-Plus sources challenging, reasoning-intensive queries from OpenAI's BrowseComp, but evaluates systems against a fixed, curated corpus of approximately 100K web documents rather than the live web.

For MAST 2026, this corpus serves as the shared English evidence collection for the multilingual and Indic query tracks.

How to Use This Dataset

Evaluating end-to-end agents: Use the MAST query sets together with this corpus to test combinations of retrievers and LLM agents. Systems receive multilingual queries, retrieve evidence from this English corpus, and generate English answers.

Evaluating retriever-only systems: Use retrieved document IDs against the human relevance labels from BrowseComp-Plus / MAST query releases to compute retrieval metrics such as Recall and nDCG.

Dataset Structure

Each example in this corpus contains:

• —docid: document identifier.
• —text: full document text.
• —url: source URL for the document.

The corpus files are stored as Parquet shards under data/.

Dataset Statistics

• —Documents: 100,195
• —Source benchmark queries: 830
• —Per-query averages in BrowseComp-Plus: 6.1 evidence docs, 76.28 negatives, 2.9 gold docs
• —Document length: 5,179.2 words on average, 32,296.2 characters on average

Related MAST Query Sets

• —MAST Multilingual Queries 2026: https://huggingface.co/datasets/mast-benchmark/multilingual-queries-2026
• —MAST Indic Queries 2026: https://huggingface.co/datasets/mast-benchmark/indic-queries-2026

Related Links

• —MAST website: https://mast-benchmark.github.io/
• —Source corpus: https://huggingface.co/datasets/Tevatron/browsecomp-plus-corpus
• —BrowseComp-Plus GitHub repository: https://github.com/texttron/BrowseComp-Plus/

Citation

@article{chen2025BrowseCompPlus,
      title={BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent}, 
      author={Chen, Zijian and Ma, Xueguang and Zhuang, Shengyao and Nie, Ping and Zou, Kai and Liu, Andrew and Green, Joshua and Patel, Kshama and Meng, Ruoxi and Su, Mingyi and Sharifymoghaddam, Sahel and Li, Yanxi and Hong, Haoran and Shi, Xinyu and Liu, Xuye and Thakur, Nandan and Zhang, Crystina and Gao, Luyu and Chen, Wenhu and Lin, Jimmy},
      year={2025},
      journal={arXiv preprint arXiv:2508.06600}
}