187,712 papers across four domains, parsed from PDF into Markdown. This is the document collection that agents search over; the benchmark queries and relevance judgments are released separately.

Domain	Papers	Shards
`computer_science`	50,048	395
`healthcare`	52,100	423
`humanities`	34,996	283
`natural_science`	50,568	408
Total	187,712	1,509

SAGE-Corpus/
├── corpus/
│   ├── computer_science/computer_science_0000.json    # 395 shards, ~124 papers each
│   ├── healthcare/healthcare1_0000.json               # 423
│   ├── humanities/humanities1_0000.json               # 283
│   └── natural_science/natural_science1_0000.json     # 408
└── metadata/
    ├── computer_science_mapping.json   # 82 MB
    ├── healthcare_mapping.json         # 109 MB
    ├── humanities_mapping.json         # 66 MB
    └── natural_science_mapping.json    # 100 MB

A shard's filename is exactly the hf_batch value in metadata/, so a paper's record points straight at the file holding it. Pull one domain with

snapshot_download("HughieHu/SAGE-Corpus", repo_type="dataset",
                  allow_patterns="corpus/healthcare/*")

Schema

Corpus shards

Each shard is a JSON list of paper records:

[
  {
    "id": "doc_aac54cd179",
    "url": "",
    "markdown": "---\nid: doc_aac54cd179\nsource_url: https://...\ntype: ...\n---\n\n# Title\n\n..."
  }
]

markdown holds the full parsed text with a YAML front-matter header. The top-level url field is usually empty — use metadata/ for provenance, joining on id ↔ doc_id.

Parsing is PDF-to-Markdown and imperfect: tables, equations and figure captions are degraded, and a small number of documents are truncated or failed to parse.

`metadata/*_mapping.json`

JSON list, one record per paper, same schema in all four domains:

Field	Type
`doc_id`	str	joins to the shard record's `id`
`hf_batch`	str	the shard holding this paper — `corpus/<domain>/<hf_batch>.json`
`hf_index`	int	its position within that shard
`url`	str	original source URL
`paper_title`	str
`domain`	str
`authors` `year` `venue` `citation_count` `abstract`	list / int / str / int / str	bibliographic fields
`metadata_source`	str	how the bibliographic fields were obtained

Identity and provenance (doc_id, hf_batch, hf_index, url, domain) are complete for every paper. The bibliographic fields are resolved against Semantic Scholar and are not: coverage is 96% for authors/year, 95% for venue and 90% for abstract, and a paper that could not be matched confidently is left blank rather than filled with a near-miss. metadata_source records the route — an identifier taken from the URL, a DOI printed in the paper's own text, or semantic_scholar:title, which was matched on title alone and is the weakest of them; filter those out if you need stricter provenance.

Loading

import json
from huggingface_hub import hf_hub_download, list_repo_files

REPO = "HughieHu/SAGE-Corpus"

# one shard
p = hf_hub_download(REPO, "corpus/computer_science/computer_science_0000.json",
                    repo_type="dataset")
papers = json.load(open(p))          # list of {id, url, markdown}

# metadata, joined on doc_id
m = hf_hub_download(REPO, "metadata/computer_science_mapping.json", repo_type="dataset")
meta = {e["doc_id"]: e for e in json.load(open(m))}
print(meta[papers[0]["id"]]["paper_title"])

To pull a whole domain:

files = [f for f in list_repo_files(REPO, repo_type="dataset")
         if f.startswith("corpus/healthcare/")]

Retrieval indices

Indices built over this corpus (256-word chunking, 4,937,036 chunks) are released separately. Each comes in a raw/ variant and an augmented/ variant, the latter encoded with the paper's corpus augmentation — domain terms extracted per chunk by Qwen3-Next-80B-A3B-Instruct, appended with the paper's title and venue.

Repository	Contents	Size
SAGE-Index-BM25	BM25 indices + the chunk metadata shared by every retriever	23 GB
SAGE-Index-Qwen3	Qwen3-Embedding-8B, `raw` + `augmented`	162 GB
SAGE-Index-ReasonIR	ReasonIR, `raw` + `augmented`	162 GB
SAGE-Index-Qwen3-LoRA	Qwen3-Embedding-8B + LoRA, `augmented`	81 GB

Start with SAGE-Index-BM25: it needs no GPU and carries the chunk_meta/ that the dense indices index into.

Sources

3,279 distinct source hosts. The 12 largest:

Host	Papers	Share
arxiv.org	41,020	21.9%
mdpi.com	24,970	13.3%
frontiersin.org	14,984	8.0%
doi.org	14,482	7.7%
nature.com	13,703	7.3%
link.springer.com	8,597	4.6%
ncbi.nlm.nih.gov	6,875	3.7%
europepmc.org	6,272	3.3%
aclanthology.org	4,593	2.5%
openaccess.thecvf.com	4,430	2.4%
cell.com	3,851	2.1%
journals.plos.org	2,483	1.3%

Roughly 64% come from hosts that are unambiguously open access (arXiv, MDPI, Frontiers, ACL Anthology, CVF Open Access, PLOS, PMC/Europe PMC, BMC).

Licensing and copyright

The compilation — the selection, parsing, sharding and metadata in this repository — is released under CC-BY-NC-4.0.

The individual papers retain the copyright of their original publishers and authors. They are reproduced here as parsed text for non-commercial scientific research, specifically the evaluation of retrieval systems. Every record carries its url in metadata/, so any document can be traced to its source of record.

If you are a rights holder and want a document removed, open an issue on this repository and it will be taken out.

Users are responsible for complying with the terms of the original publishers when redistributing or building on this data.

Citation

@article{sage2026,
  title  = {SAGE: Benchmarking and Improving Scientific Literature Retrieval for Deep Research Agents},
  author = {TBD},
  year   = {2026}
}

Downloads last month: 81

Total file size:

9.47 GB

SAGE-Corpus

Contents