DriftCVE-CWE v3.0.0

DriftCVE-CWE is a chronologically split benchmark for classifying CVE vulnerability descriptions into CWE weakness categories, built to measure how text classifiers trained on past vulnerabilities behave on later ones. It accompanies the working paper DriftCVE-NLP: Evaluating Vulnerability Retrieval and Classification Under Temporal Language Drift (Arun Kumar Gharami, Florida Atlantic University, CAP 6640) and the code at https://github.com/Arungharami/DriftCVE-NLP.

Version: 3.0.0 (schema 3.0) · Collection cutoff: 2026-09-09T23:59:59.999000+00:00 · Raw snapshot SHA-256: 88bb70a8d3f3837713f41e5be7c7af1e53c4d7bb0c85b5f32efa22526ac6bbfb

This product uses the NVD API but is not endorsed or certified by the NVD. Records were filtered, de-duplicated and assigned to splits by DriftCVE-NLP; the selection is not an NVD product.

Loading

from datasets import load_dataset

ds = load_dataset("arun-gharami/driftcve-cwe", revision="11fdd5179613b029a4fb418a9d595bba2073e84d")   # pinned commit for v3.0.0
test_2025 = ds["test"]
future_2026 = ds["future"]                           # partial year, see cutoff above
# Records with a single specific CWE outside the 20 training labels (open-set analysis):
outside = load_dataset("arun-gharami/driftcve-cwe", "out_of_label_set", revision="11fdd5179613b029a4fb418a9d595bba2073e84d")["records"]

Always pin revision. Newer versions are added as new commits and tags; existing revisions are never rewritten. The previous CSV release is still available as tag v2.0.0 (commit 2a980fd9057e1c0c18818401259b839616d7a201).

Purpose and intended use

• Research and teaching on temporal generalization of CVE→CWE text classification and vulnerability retrieval.
• Comparing classical, embedding and transformer classifiers under a fixed chronological split.

Not intended for: automated triage or labeling decisions without expert review, judging the security of particular products or vendors, or any claim that the labels are ground truth (see Limitations).

Source and provenance

• Source: NVD CVE API 2.0 (services.nvd.nist.gov/rest/json/cves/2.0), queried by publication date windows from ['2018-01-01', '2026-09-09'].
• Snapshot policy: Historical cache plus overlapping cutoff extension; not a single-time NVD snapshot. Lineage (file checksums and collection timestamps) is recorded in dataset_manifest.json.
• Descriptions are the English NVD descriptions as returned by the API; CWE labels are the weakness entries attached to each record at collection time (NVD and CNA sources merged).
• last_modified_at reflects the record state at collection, not at publication. Labels of old records may have been revised after publication; historical experiments therefore use current metadata, not the information available at the time.

License and terms

No new license is granted by this repository; use is governed by the source terms:

• CVE records — CVE Program Terms of Use (https://www.cve.org/Legal/TermsOfUse): MITRE grants a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare derivative works of, publicly display, publicly perform, sublicense, and distribute CVE, provided that MITRE's copyright designation and this license are reproduced in any copy. Provided "AS IS" without warranties.
• NVD API — NVD API Terms of Use (https://nvd.nist.gov/developers/terms-of-use): the notice above is displayed; the NVD name identifies the source only and implies no endorsement; modified content must not be attributed to the NVD.
• CWE identifiers — CWE Terms of Use (https://cwe.mitre.org/about/termsofuse.html), © The MITRE Corporation; CWE is a trademark of The MITRE Corporation.

Copyright © The MITRE Corporation for CVE and CWE content. The derived split assignment, manifest and card are provided under the same terms.

Schema

column	type	description
cve_id	string	CVE identifier
description	string	English NVD description (verbatim)
published_at	timestamp (UTC)	NVD publication time; determines year and split
last_modified_at	timestamp (UTC)	NVD last-modified time at collection
year	int	publication year
split	string	train / validation / test / future
cwe_labels	list[string]	all weakness entries on the record, sorted
selected_label	string	the single specific CWE used as the target
label_id	int	index of selected_label in the numeric-sorted label list
source_url	string	NVD detail page
collection_cutoff	string	collection cutoff date
cvss_version, cvss_score, severity	string/float/string	primary CVSS metadata (not a model input)
cpe_vendors	list[string]	vendors from CPE configurations (shortcut analysis only)
mentions_cwe_id	bool	description contains a literal CWE-<n> string
nearest_train_cosine	float	max TF-IDF cosine to any training description (null in train)

Splits

Chronological by published_at; every split is strictly after the previous one.

split	years	rows	first published	last published	partial year
train	2018, 2019, 2020, 2021, 2022, 2023	63,521	2018-01-01	2023-12-31	no
validation	2024	21,492	2024-01-01	2024-12-31	no
test	2025	21,103	2025-01-01	2025-12-31	no
future	2026	23,321	2026-01-01	2026-09-09	yes

The future split is a partial year (1 January 2026 to the collection cutoff).

Label policy

Keep records with exactly one specific CWE- weakness after removing NVD-CWE-noinfo/NVD-CWE-Other placeholders; the label set is the top-K most frequent single labels in training years only (ties broken by CWE id). Label selection used years 2018–2023 only. Exclusions are applied in this order:

step	rule	removed	remaining
0	input_records	0	290,522
1	rejected_or_withdrawn	13,932	276,590
2	missing_description	0	276,590
3	missing_cwe	5,559	271,031
4	generic_nvd_cwe_noinfo	18,142	252,889
5	generic_nvd_cwe_other	1,789	251,100
6	multiple_specific_cwes	32,143	218,957
7	published_outside_protocol_window	0	218,957
8	duplicate_cve_id	0	218,957
9	duplicate_normalized_description	10,923	208,034
10	year_outside_split_protocol	0	208,034
11	outside_training_top_k_label_set	78,597	129,437

Notes: rejected or withdrawn records are detected from the NVD description prefix (** REJECT ** or Rejected reason:). Exact duplicates are detected after NFKC normalization, case folding and whitespace collapsing, and the earliest-published copy is kept, so a later split never repeats an earlier text verbatim. Near duplicates are not removed; they are measured and stored in nearest_train_cosine:

split	rows	nearest_train_cosine >= 0.9	share
validation	21,492	816	3.8%
test	21,103	789	3.7%
future	23,321	210	0.9%

Descriptions containing a literal CWE identifier (potential label leakage): train 402, validation 105, test 146, future 182.

Class distribution

label	train	validation	test	future
CWE-79	15,754	6,684	6,870	4,758
CWE-787	6,355	907	639	869
CWE-89	5,782	2,418	1,653	2,052
CWE-20	3,766	477	423	971
CWE-125	3,533	911	719	1,013
CWE-352	3,524	1,351	1,805	736
CWE-22	3,009	857	966	1,784
CWE-200	2,576	685	598	941
CWE-416	2,516	1,073	858	1,535
CWE-78	2,393	677	664	1,055
CWE-434	1,724	670	545	465
CWE-120	1,638	515	351	388
CWE-119	1,629	128	218	155
CWE-476	1,587	1,066	1,081	730
CWE-862	1,584	1,712	2,159	2,762
CWE-287	1,526	220	265	522
CWE-190	1,438	185	226	375
CWE-269	1,092	328	291	550
CWE-400	1,076	302	329	553
CWE-863	1,019	326	443	1,107

Coverage of the training label set by year

Records with one specific CWE outside the 20 training labels are excluded from the benchmark and released separately in the out_of_label_set configuration (78,597 rows).

year	split	single-label records	in label set	outside label set	coverage
2018	train	12,852	9,766	3,086	76.0%
2019	train	12,974	8,799	4,175	67.8%
2020	train	12,520	8,028	4,492	64.1%
2021	train	14,460	9,337	5,123	64.6%
2022	train	18,763	12,590	6,173	67.1%
2023	train	22,018	15,001	7,017	68.1%
2024	validation	32,431	21,492	10,939	66.3%
2025	test	35,840	21,103	14,737	58.9%
2026	future	46,176	23,321	22,855	50.5%

Limitations

• NVD weakness assignments are noisy, change over time and mix NVD and CNA judgments; they are not expert ground truth. Multi-label records are excluded, which biases the benchmark toward simpler weaknesses.
• Coverage of the fixed training label set falls in later years, so results on 2025–2026 describe only the in-set portion of those years.
• The 2026 split is partial; NVD weakness assignment can lag publication, so its composition may change as records are analyzed.
• The snapshot combines an initial collection and an overlapping extension rather than a single instant; record metadata may have changed since.
• Descriptions contain vendor and product names, which can act as shortcuts.

Revisions

tag	change
v2.0.0	Single CSV release (DriftCVE-CWE-v2), no named splits.
v3.0.0	Parquet splits, manifest and card. The out_of_label_set config used the reserved split name all and cannot be loaded at this tag; the benchmark config loads normally.
v3.0.1	Card/config fix only: out_of_label_set split renamed to records. Data files are byte-identical to v3.0.0 (checksums below).

Files

path	sha256
data/benchmark/train.parquet	43a9838e7a933d42bf6d4036accacb3caa75b73fd75f4588ce560f4f5b377fbd
data/benchmark/validation.parquet	7a4e1823fb1c4ce65f87bdd1870ee2cb0830a688ded21f22be39897404ad198b
data/benchmark/test.parquet	8bb763559f18ff91c9d4d4eb560383e533832f0b3a34e35e5e1c7108525cef40
data/benchmark/future.parquet	7cb3933ecfd7bfeeee322138fc32d2e244c21a2ae96e6dd85409f1b5ab6f62d6
data/out_of_label_set/all.parquet	99968f859e276a7e76ae3bfdce84e54b1e0d1f891ebf0fb7412e80752204c6fb

Citation

@misc{gharami2026driftcve,
  author = {Gharami, Arun Kumar},
  title  = {DriftCVE-NLP: Evaluating Vulnerability Retrieval and Classification Under Temporal Language Drift},
  year   = {2026},
  note   = {Working paper, CAP 6640 Natural Language Processing, Florida Atlantic University.
            Dataset DriftCVE-CWE v3.0.0},
  howpublished = {\url{https://huggingface.co/datasets/arun-gharami/driftcve-cwe}}
}