Request access to VirBench
VirBench is an expert-curated benchmark of 120 viral sequence retrieval queries with manually established reference counts and reference accession sets. Access is gated so that the query-reference pairs do not enter language model training corpora, which would allow a model to emit a correct count without performing retrieval. Requests are reviewed by the authors, normally within five working days.
By requesting access you agree not to republish the reference values and not to post them anywhere they can be crawled, including in public repositories, supplementary files, or model training data.
Log in or Sign Up to review the conditions and access this dataset content.
VirBench
VirBench is an expert-curated benchmark of 120 viral sequence retrieval queries with manually established reference counts and reference accession sets. It accompanies Deterministic access to global viral sequence data enables robust agentic scientific discovery (Nasri et al., 2026).
Each query specifies a pathogen, by NCBI TaxID or accession, together with a combination of metadata filters drawn from real virology workflows: host, geographic location, collection and release date ranges, sequence length bounds, nucleotide completeness, source database, genome segment, lineage, submitter country and vaccine strain status. Queries span 40 pathogens across multiple taxonomic levels, from searches over all viruses or entire families down to species-level queries and individual accession lookups. Each query applies between 1 and 9 filters simultaneously, with a median of 6. Fifty-eight queries were contributed by the Sabeti Lab diagnostics team from recurring assay development and surveillance workflows.
Reference values were established by manual execution in the NCBI Virus web interface. Expected counts range from 0 to 3,226 with a median of 21.5, and sum to 26,752 records across the benchmark.
Why this deposit is gated
Publishing the query-reference pairs where a language model can encounter them would turn the benchmark into an answer key. A model could then emit a correct count without performing retrieval, and the benchmark would measure recall of a published answer rather than retrieval ability.
Access is granted on request to any researcher, under a condition of use: recipients agree not to republish the reference values and not to post them anywhere they can be crawled.
Everything that does not carry reference values is public without restriction,
including the benchmark harness, the build_query() prompt-construction
function, the exact system and user prompts, the agent-facing tool
documentation, the scoring and analysis code, the curation protocol and the
aggregate per-arm results:
https://github.com/lauraluebbert/VirBench
The analysis pipeline can therefore be inspected and re-executed end to end on any query set without access to this deposit.
Contents
| Path | Description |
|---|---|
benchmark/virbench_v1.csv |
Reference values, version 1 |
benchmark/virbench_v2.csv |
Reference values, version 2 |
benchmark/virbench_v3.csv |
Reference values, version 3 |
benchmark/virbench_v4.csv |
Reference values, version 4. Current version. Adds per-query pointers to the reference accession sets and to the corresponding NCBI Virus web query |
reference_sets_v4.tar.gz |
Complete record tables exported from the NCBI Virus web interface, one file per query. |
VERSIONS.md |
Version history, what changed between versions, and which evaluation arm was scored against which version |
USAGE.md |
Conditions of use, citation, and guidance on reference value drift |
Each query CSV holds 120 rows, one per query, keyed by query_id. The
expected_count column carries the reference value; the remaining columns
define the query.
Extract the archive with tar xzf reference_sets_v4.tar.gz. It unpacks to
reference_sets/v4/, with filenames matching the
ncbi_virus_web_result_file_name column of virbench_v4.csv.
Versions
Reference values are versioned because records are on occasion deposited into INSDC retrospectively, or withdrawn or reclassified after deposition. Every query except the three accession lookups is bounded by a maximum release date, so the matching record set is closed and each reference value is fixed rather than drifting with database growth. That bound is strong but not absolute, which is what the versioning records.
| Version | Curated | Reference values changed | Arms scored against it |
|---|---|---|---|
| v1 | February 2026 | Initial curation | Claude Sonnet 4, Biomni OSS, Edison Analysis, GPT-5.2-pro, standalone gget virus |
| v2 | May 2026 | 9 of 120 differ from v1 | Claude Opus 4.7, GPT-5.5 |
| v3 | 2026 | 5 of 120 differ from v2 | None; superseded before use in the reported evaluation |
| v4 | September 2026 | 2 of 120 differ from v3. 5 differ from v1 and 7 from v2. Adds reference accession sets | Standalone gget virus, re-executed for the set-level comparison |
Date fields are
written in different formats across versions, and are reproduced here exactly as
executed rather than normalised. See VERSIONS.md.
Access
Click Request access above and provide your name, affiliation and intended use. Requests are reviewed by the authors.
Citation
Nasri, F., Gurev, S., Varilly, P., Ramesh, K., O'Leary, N. A., Cool, J., Renard, B. Y., Sabeti, P. C., and Luebbert, L. (2026). Deterministic access to global viral sequence data enables robust agentic scientific discovery. arXiv:2606.06749. https://arxiv.org/abs/2606.06749
Contact
- Downloads last month
- 29