VulSpot

VulSpot is a statement-level dataset for vulnerability detection, CWE classification, and localization. It follows the SecVulEval collection pipeline (NVD CVEs with a Patch tag, GitHub fixing commits, function extraction, filtering) through 2026, and adds AST-level statement diffs, sink-candidate anchors for additive fixes, label-confidence tiers, untouched negatives, and deterministic local context. It does not include the LLM-generated §3.4 context of SecVulEval.

from datasets import load_dataset

cpp = load_dataset("DynaOuchebara/VulSpot", "cpp", split="full")
multi = load_dataset("DynaOuchebara/VulSpot", "multi", split="full")

Configs

config	contents
cpp	C/C++ only (.c/.cc/.cpp/.h/.hpp/...)
multi	every supported language, with a language column

Each config has two splits:

• full — every row that survived structural filters
• balanced — full with a per-year cap on post-2024 kernel.org-CNA Linux CVEs (see is_kernel_cna)

Chronological and project-holdout index files live next to the parquet (splits/chrono.json, splits/project_holdout.json). They refer to the idx column.

Collection

1. NVD (API 2.0): all non-rejected CVEs, windowed by publication date, incremental via lastModified.
2. Patch commits: references tagged Patch whose URL is a GitHub commit (plus git.kernel.org remapped to torvalds/linux). Other hosts are counted in the pipeline report and not fetched. Forks are kept only if the SHA exists on the parent, and are attributed to the parent.
3. Commit artifacts: GitHub REST + raw.githubusercontent.com (message, unified diff, before/after file contents). Merge commits are dropped.
4. Functions: tree-sitter extraction; before/after pairing by name + signature + order; statement-unit AST diff.
5. Filters: extension filter; multi-commit CVEs dropped; files with no in-function edits dropped; tangled-change heuristic becomes a confidence tier rather than a hard drop (unmentioned functions in multi-function files are still dropped).

Label semantics

Two localization layers:

• changed_lines / changed_statements (fix_*): SecVulEval semantics. On a vulnerable row these are the deleted/updated statements; on a fixed row, the added/updated ones. Serialized as JSON strings [[line, text], ...] with file-absolute line numbers so existing SecVulEval loaders work.
• sink_candidates (vulnerable rows only): deleted + updated statements, plus insertion anchors in the vulnerable version (the next statement, or, for a guard if (...) return/goto/break/continue, the first later use of a guard identifier). Each entry has abs/rel line, byte range, text, and kind.

is_vulnerable is the function-level flag. cwe_list / cwe_1003_list are the type labels (raw NVD CWEs and the CWE-1003 simplified-mapping view).

Confidence tiers (label_confidence)

• high — the commit changes exactly one function (test files excluded), or the function is named in a linked CVE description
• medium — the only changed function in its file, but the commit touches several files
• low — several functions changed in the file; this one is named in the commit message

Negatives (neg_type)

• fixed — patched version of a vulnerable function
• untouched — unchanged function from a file that also contains a changed function
• patched_later_vulnerable — a fixed row whose function is modified again by a later CVE commit in this dataset (later_fixed_by)

clean is true when label_confidence ∈ {high, medium} (or the row is an untouched negative), label_conflict is false, and neg_type is not patched_later_vulnerable.

is_kernel_cna flags CVEs assigned by the kernel.org CNA with published_date ≥ 2024-02-01.

Splits

• splits/chrono.json — 80/10/10 by CVE published_date, then clone-group aware (clone_group_id): a row whose near-duplicate group already appears in an earlier split (train < val < test) is excluded, so train never contains rows newer than test.
• splits/project_holdout.json — a random ~10% of rows, taken as whole projects, as test (Linux always stays in train/val); the rest split chronologically into train/val. Same clone-group exclusion.

Both files also contain evaluation subsets of test:

• test_high — clean test rows, without vulnerable rows whose label_confidence is not high (all negatives kept)
• test_paired_high — high-confidence vulnerable rows whose fixed version is also in test, plus those fixed versions (PrimeVul-style paired evaluation)
• test_no_linux — test without torvalds/linux rows

In balanced, the post-2024 kernel CVEs kept each year are a seeded sample spread over the year, so chronological val/test still contain some Linux.

Exact duplicates use hash (md5 of per-line-stripped body). norm_hash drops comments/whitespace. clone_group_id is MinHash-LSH (5-shingles, Jaccard ≥ 0.8).

Licensing

Function bodies are reproduced verbatim from the original open-source repositories for research. Each row points back to its source via project_url and commit_id. The code remains under the license of its original project. This dataset does not grant or claim a single license covering the code.