Claim · 2024-05-15 · publication · Stated in source · Supported

The WMDP benchmark paper states that it publicly releases a dataset of 3,668 multiple-choice questions serving as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity and chemical security, developed by a consortium of academics and technical consultants and filtered to remove sensitive information before release.

One recorded claim. The claim panel gives its qualification, the source passages with their place in each source, and how it was checked.

What it connects

The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning (arXiv:2403.03218) → reports evaluation of →

Where it appears

No collection page holds this claim. It is listed with its records in the explorer.

Claim and evidence

The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning (arXiv:2403.03218)reports evaluation of ↓

Machine-checked, not reviewed by an editorStated in sourceSupported

Supporting evidence: primary · 1 source

Subject version: Version 7 of 15 May 2024, as described on the arXiv abstract page · Published 2024-03-05 · Effective date not established · Revised 2024-05-15
Date: 2024-05-15 · publication

Scope and limitsRecords what the benchmark states it measures and how its authors describe its release, not any finding about a system. The paper's unlearning method and results are deliberately not recorded. Only the abstract page was read. The authors' statement that the dataset was filtered is their own. Only the benchmark's stated object of measurement and release statement are recorded; the paper's method sections were not read.

The evidence 02

primary sourceSupports the claim

arXiv abstract page for arXiv:2403.03218 (The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning)

arXiv (Cornell University) · Publication date unknown

Cite the original source

Cite the original document, not Selekoa. Check these details against it.

Source excerpt

“a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security”

Location in source Abstract, fourth sentence

How it was read Retrieved directly from the publisher

Machine-checked · 2026-10-30

Source details and check notes

Authors: arXiv (Cornell University)
Version: Retrieved for the candidate pack on 2026-10-28 (SHA-256 c826f5d118daa2d0d9fda28de81b4460006387671abeb09faf704befe21c1c3e). Abstract page as served on 2026-10-28, listing v1 of 5 March 2024 and v7 of 15 May 2024 (the current version). Only the abstract page was read; the paper's unlearning method is deliberately not recorded. Retrieved again at admission on 2026-10-30 (SHA-256 111ae33a8fe87cffd7c31c9ba63f30ff31cb893d6f1cb86cb94767084668b500). Its wording can change; earlier wording is not authenticated.
Canonical address: Open the canonical source ↗
Accessed: 2026-10-28
Language: English (en)
Attribution: arXiv (Cornell University), “arXiv abstract page for arXiv:2403.03218 (The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning)”.
Reuse: No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution.
Check notes: Candidate pack passage, AI and biosecurity (research/ai-bio). The quotation (21 words) matched the text retrieved for the candidate pack on 2026-10-28 and again at admission on 2026-10-30, in a fresh retrieval (SHA-256 111ae33a8fe87cffd7c31c9ba63f30ff31cb893d6f1cb86cb94767084668b500; the bytes differ from the pack’s retrieval). Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. No human editorial review.

primary sourceSupports the claim

arXiv abstract page for arXiv:2403.03218 (The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning)

arXiv (Cornell University) · Publication date unknown

Cite the original source

Cite the original document, not Selekoa. Check these details against it.

Source excerpt

“was stringently filtered to eliminate sensitive information prior to public release”

Location in source Abstract, fifth sentence

How it was read Retrieved directly from the publisher

Machine-checked · 2026-10-30

Source details and check notes

Authors: arXiv (Cornell University)
Version: Retrieved for the candidate pack on 2026-10-28 (SHA-256 c826f5d118daa2d0d9fda28de81b4460006387671abeb09faf704befe21c1c3e). Abstract page as served on 2026-10-28, listing v1 of 5 March 2024 and v7 of 15 May 2024 (the current version). Only the abstract page was read; the paper's unlearning method is deliberately not recorded. Retrieved again at admission on 2026-10-30 (SHA-256 111ae33a8fe87cffd7c31c9ba63f30ff31cb893d6f1cb86cb94767084668b500). Its wording can change; earlier wording is not authenticated.
Canonical address: Open the canonical source ↗
Accessed: 2026-10-28
Language: English (en)
Attribution: arXiv (Cornell University), “arXiv abstract page for arXiv:2403.03218 (The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning)”.
Reuse: No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution.
Check notes: Candidate pack passage, AI and biosecurity (research/ai-bio). The quotation (11 words) matched the text retrieved for the candidate pack on 2026-10-28 and again at admission on 2026-10-30, in a fresh retrieval (SHA-256 111ae33a8fe87cffd7c31c9ba63f30ff31cb893d6f1cb86cb94767084668b500; the bytes differ from the pack’s retrieval). Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. The authors' own statement about the released dataset. No human editorial review.

Status and meaning

Editorial status. Not reviewed by an editor. An exploratory record that no editor has accepted. This status does not say whether a person has read any linked passage. Each evidence record above shows how it was checked.

Relation. A document names a concept as an evaluation category. This is not a finding that the evaluated system exhibits the property.

Date. Version 7 of the preprint, as described on the arXiv abstract page; version 1 was submitted on 5 March 2024.

Record ID ab-ev-wmdp-benchmark-measures