YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SALVE Dataset — Schema & Datasheet
Formal documentation of the curated secure/insecure JavaScript & TypeScript vulnerability dataset (T1 deliverable "Dataset Documentation & Benchmarking Guidelines"). Follows the Datasheets for Datasets structure (Gebru et al.).
This datasheet describes the active, versioned dataset
salve-core-le9-v2 (built 2026-06-18). All figures below are measured from
that file, not estimated. The dataset is resolved through the registry
(data/datasets/registry.json); see §8 for versioning. The raw, pre-clean
collection pool is data/js_vuln_pairs.jsonl (§3).
Companion docs: collection methodology →
data_collection.md; quality control →preprocessing.md§1; file-count noise →file_count_investigation.md; CWE/Joern validation →cwe_frequency_and_joern.md,oracle_validation.md.
1. Overview / Motivation
Pairs of vulnerable (pre-patch) and fixed (post-patch) real-world JS/TS source files, each drawn from a security-fix commit and labelled with a CWE type. Built to (T1) study security-preserving perturbations and (T2) measure whether LLM vulnerability classifiers are robust to those perturbations. Real fixes from real projects — not synthetic — so the secure/insecure distinction reflects how vulnerabilities are actually introduced and patched.
2. Composition
2.1 Size (measured — salve-core-le9-v2)
| Granularity | Count |
|---|---|
| Records (one per changed core file) | 7,075 |
Unique vulnerabilities (vuln_id) |
3,727 |
| Unique projects | 1,328 |
| Unique fix commits | 2,310 |
| Complete pre+post pairs | 5,184 (73.3%) |
| Malformed JSON | 0 |
| Language split | JS 58.8% / TS 41.2% |
How this relates to the collection pool. The raw extraction
(js_vuln_pairs.jsonl) yields 3,943 JS/TS single-CWE single-commit vulns
across 16,251 deduped file-records. The active dataset applies quality control
(§4) — dedup, core-only, a ≤9 core-files-per-commit cut, and a min-patch
filter — retaining 3,727 / 3,943 = 94.5% of the collected vulnerabilities
while discarding release/refactor noise. (See data_collection.md for the full
upstream→collected provenance funnel: upstream 44,037 all-language vulns →
single-CWE → single-commit → JS/TS → 3,943 collected.)
Read this carefully: one row = one file in a fix commit, not one vulnerability. A vulnerability may span several files (now at most 9, by construction), and a fix commit may carry several vulnerabilities.
2.2 Record schema (one JSON object per line)
| Field | Type | Completeness | Description |
|---|---|---|---|
id |
str | 100% | Composite primary key: vuln_id::commit_sha::file_path. Dedup key. |
cwe |
str | 100% | CWE identifier, e.g. CWE-79. Exactly one CWE per vuln (0 conflicts). |
cve |
str/null | 55.1% | CVE id when one exists. Null for the ~45% that are GHSA-only advisories. Use metadata.vuln_id as the universal key. |
language |
str | 100% | javascript or typescript. |
pre_patch |
str | 77.6% non-empty | Removed (vulnerable) lines of the unified diff, - stripped — not the whole file. Empty when the fix only adds lines. Full pre-fix files → data/reconstructed/ (§2.7). |
post_patch |
str | 95.6% non-empty | Added (fixed) lines of the unified diff, + stripped. Empty when the fix only deletes lines. |
commit_url |
str | 100% | GitHub commit URL for the fix. |
metadata.vuln_id |
str | 100% | Advisory ID — CVE-… or GHSA-…. The de-facto identifier. |
metadata.summary |
str | 100% | Advisory summary text. |
metadata.score |
num | 95.7% | CVSS base score. |
metadata.severity |
str | 95.7% | Qualitative band (none/low/medium/high/critical), derived from score. |
metadata.published_date |
str | 100% | ISO date of the advisory. |
metadata.project |
str | 100% | Source repository URL. |
metadata.commit_sha |
str | 100% | Fix commit SHA (parent = sha~1 = the vulnerable state). |
metadata.file_path |
str | 100% | Path of the file within the repo. |
metadata.additions |
int | 100% | Lines added in this file's diff (median 6, mean 22, max 2,293). |
metadata.deletions |
int | 100% | Lines deleted (median 2, max 2,454). |
metadata.shared_commit |
bool | 100% | true if the file shares its commit with other vuln files. |
metadata.category |
str | 100% | File category from quality control — always "core" in this dataset (§4 Pass 1). |
metadata.relevance_score |
float | optional | Present only when the relevance filter (Pass 4) scored the file; absent in v2 (relevance filter off). |
2.3 CWE composition — by unique vulnerability (canonical)
Top-12 of 3,727 vulns (one CWE each); top-10 = 62.7%, top-12 = 66.5%.
| Rank | CWE | Name | Vulns | % |
|---|---|---|---|---|
| 1 | CWE-79 | Cross-site Scripting (XSS) | 1,131 | 30.3% |
| 2 | CWE-1321 | Prototype Pollution | 231 | 6.2% |
| 3 | CWE-22 | Path Traversal | 192 | 5.2% |
| 4 | CWE-918 | SSRF | 140 | 3.8% |
| 5 | CWE-78 | OS Command Injection | 131 | 3.5% |
| 6 | CWE-352 | CSRF | 110 | 3.0% |
| 7 | CWE-200 | Information Exposure | 108 | 2.9% |
| 8 | CWE-94 | Code Injection | 108 | 2.9% |
| 9 | CWE-1333 | ReDoS | 95 | 2.5% |
| 10 | CWE-20 | Improper Input Validation | 89 | 2.4% |
| 11 | CWE-400 | Uncontrolled Resource Consumption | 81 | 2.2% |
| 12 | CWE-601 | Open Redirect | 64 | 1.7% |
2.4 CWE composition — by file-record
Counted by record (7,075): CWE-79 30.0%, CWE-22 5.1%, CWE-1321 5.0%, CWE-918 4.7%, CWE-352 3.5%, CWE-200 3.1%. The by-vuln and by-file rankings now agree (both led by XSS at ~30%). This is a side effect of the ≤9-files-per-commit cut: the old corpus showed a large CWE-95 (Eval Injection) spike at file level (13.6%) driven by a few vulns whose fix commits touched hundreds of files — those release/refactor commits are exactly what the commit cut removes, so the file-level skew is gone.
2.5 Pair completeness (critical for paired use)
| Side | Count | % |
|---|---|---|
| Complete pair (pre and post non-empty) | 5,184 | 73.3% |
Empty pre_patch (file added by fix) |
1,584 | 22.4% |
Empty post_patch (file removed by fix) |
309 | 4.4% |
For any task needing both a vulnerable and a fixed version of the same file (secure/insecure classification, perturbation pairing), use the 5,184 complete pairs. The remainder are single-sided by nature (vendored additions, deletions). Note these are diff hunks, not full files — for whole-file analysis use the reconstructed corpus (§2.7). Completeness improved markedly over the collection pool (59.6% → 73.3%) because the commit cut removes large new-file dumps (empty-pre records).
2.6 Language
JavaScript 4,162 (58.8%), TypeScript 2,913 (41.2%). Both first-class; TS
matters for cross-file taint resolution (see oracle_validation.md).
2.7 Reconstructed full files (data/reconstructed/)
extraction/reconstruct_files.py fetches the complete pre- and post-fix
files from GitHub (post = file at the patch commit; pre = file at the parent
commit, with reverse-diff fallback) and writes them as
data/reconstructed/<CWE>/<vuln_id>__<safe_path>.{pre,post}.<ext>. This is the
corpus the Joern/Semgrep oracles run on (whole-file analysis needs full context,
not diff hunks). Currently ~957 complete full-file pairs across the top-10 CWEs
(a subset, expanding). Note: the reconstructed corpus is not yet restricted
to the active dataset's commit cut — that wiring is pending.
3. Collection process (summary)
Vulnerabilities come from a dashboard export of advisory data (OSV / GitHub
Security Advisories). extraction/fetch_dataset.py keeps only single-commit,
single-CWE advisories, takes each changed core .js/.ts/.jsx/.tsx/.mjs/.cjs
file's unified diff, and emits removed lines as pre_patch and added lines as
post_patch → data/js_vuln_pairs.jsonl (the collection pool).
extraction/reconstruct_files.py separately fetches the whole pre/post files
(§2.7). Full sourcing methodology and the upstream→collected funnel →
data_collection.md.
4. Preprocessing / quality control
preprocessing/noise_reduction.py applies, in order: (0) dedup (drops 75.3%
exact-duplicate rows from a source-CSV bug), (1) category verification (keeps
only core — drops test/docs/config/lockfile/vendored/generated), (1b) commit
core-file cap (discards whole commits touching > N core files), (2/3)
CWE/language filters (opt-in), (4) relevance filter (opt-in, signal-
preserving), (5) per-vuln file cap (opt-in), (6) minimum patch size.
The active dataset salve-core-le9-v2 is built with Pass 1b at N=9 and
min-patch ≥1 (relevance filter and per-vuln cap off):
python3 preprocessing/noise_reduction.py \
--max-commit-core-files 9 --no-cap --min-patch-lines 1 \
--dataset-version salve-core-le9-v2
Pipeline (records): raw 65,825 → dedup 16,251 → core 14,551 →
≤9-files/commit 7,075 → min-patch≥1 7,075. The ≤9 cut is the elbow of
the commit-size/retention curve (dashboard "Commit-size cutoff" page): it discards
101 large release/refactor/merge commits (~half of all core-file records) while
retaining 94.5% of vulnerabilities, because those commits' files are mostly
non-vulnerable source churn. See file_count_investigation.md.
5. Recommended uses & benchmarking guidelines
- Secure/insecure classification: use the 5,184 complete pairs (§2.5).
Treat
pre_patch= vulnerable (label 1),post_patch= fixed (label 0). - Stratify by CWE using the by-vulnerability counts (§2.3); the long tail is steep (top-1 alone is 30%).
- Split by
vuln_idorproject, not by row — files from one vuln/commit are highly correlated; a row-level split leaks. - Use
scorefor severity (severityband is derived from it). - Ground-truth "still vulnerable?" labels for perturbed variants come from the
Joern oracle (
validation/joern/cwe_specs.json), not from this file. - Pin the dataset version (
salve-core-le9-v2) and its sha256 from the registry (§8) in any experiment for reproducibility.
6. Known limitations (foundation notes)
cveis 55.1% populated — the rest are GHSA-only advisories with no CVE by nature.metadata.vuln_idis the universal key (CVE/GHSA mix).- 73.3% are complete two-sided pairs. Empty-
pre(new files) and empty-post(deletions) are inherent to commit-derived data; paired tasks must filter to the complete subset. - Vuln attribution is commit-level. A CVE is attributed to a whole commit;
the actual fix may be in a subset of its (now ≤9) files. The commit cut and the
optional relevance filter mitigate this but do not perfectly localize the fix —
"a core file" is a changed source file, not necessarily the vulnerable file
(see
file_count_investigation.md). - Inadvertent/incidental fixes are only approximately excluded (via the single-commit + commit-cut filters); oracle-verified exclusion exists for the perturbation loop, not retroactively over the base pairs.
- CWE granularity varies (e.g. CWE-94 vs CWE-95 both code/eval injection).
- Collection is by file extension, so the dataset includes JS/TS files from
polyglot repos whose primary language is not JS — there is no clean "JS share"
denominator against the upstream set (see
data_collection.md).
7. Provenance & maintenance
- Collection pool:
data/js_vuln_pairs.jsonl(gitignored — large; regenerated byextraction/fetch_dataset.py). - Active dataset:
data/datasets/salve-core-le9-v2.jsonl(+ a byte-identical copy atdata/js_vuln_pairs_clean.jsonlfor back-compat), resolved via the registry (§8). - Reconstructed files:
data/reconstructed/<CWE>/*(pre/post full files). - Regenerate: the
noise_reduction.pycommand in §4. This datasheet should be regenerated when the active dataset version changes.
8. Dataset versions & registry
Every built dataset is recorded in data/datasets/registry.json by
preprocessing/dataset_registry.py, with the exact filters, measured counts, a
sha256 of the file and its source, and a build date. The registry names one
active version; downstream code resolves it via dataset_registry.active_path()
instead of hard-coding a path, and the dashboard exposes a version selector.
| Version | Records | Vulns | Filters | Status |
|---|---|---|---|---|
salve-core-le9-v1 |
6,800 | 3,668 | core, ≤9 files/commit, min-patch ≥2 | superseded |
salve-core-le9-v2 |
7,075 | 3,727 | core, ≤9 files/commit, min-patch ≥1 | active |
v2 lowered min-patch from ≥2 to ≥1 to recover 59 legitimate one-line fixes
(e.g. a single added escape()/validation call) that ≥2 had dropped. min-patch
counts changed lines (combined pre+post), not files. Build a new version with
noise_reduction.py --dataset-version <name>; it is registered and marked active
automatically.
- Downloads last month
- 6