--- license: cc-by-4.0 task_categories: - text-classification size_category: 10K Supersedes [SpongeBOB9684/CVE-TO-MITRE](https://huggingface.co/datasets/SpongeBOB9684/CVE-TO-MITRE) (2025 real+synthetic mix): this release is 15,000 human-reviewed CVEs only, re-labelled on ATT&CK v19.2 (697 techniques), CC-BY-4.0, with a reproducible builder. ## Dataset Description Multi-label text classification: every row is a CVE description annotated with the ordered set of MITRE ATT&CK techniques an attacker would chain to exploit it. The dataset contains 15,000 rows over 697 labels (ATT&CK Enterprise v19.2, parents and sub-techniques). Each row carries the technique ids, their human-readable names, a 697-way multi-hot vector in the stable internal label order, the CVE id(s), the EPSS score when available, and the CISA Known Exploited Vulnerabilities (KEV) flag. ## Source and Provenance This dataset uses data from the CVE Program (cve.org) and MITRE ATT&CK (attack.mitre.org). The rows are a subset of the reviewed training split of an internal mapping pipeline. Upstream of labelling, descriptions were redacted (1,712 redacted descriptions dropped) and duplicate text/label groups collapsed (557 duplicates collapsed). Every row is traceable to the internal split manifest: manifest train entry i is index-aligned with train row i, and the CVE id(s) are read from that manifest's records. ### How this dataset was produced ```mermaid flowchart LR A["cvelistV5 repo
(public CVE records)"] --> B["Collector + dedup"] B --> C["Enrichment
EPSS (FIRST) + KEV (CISA)"] C --> D["Description filter
1,712 redacted dropped"] D --> E["LLM mapping
GLM → ordered ATT&CK chains"] E --> F["Rule-based validation
id format · confidence · evidence
vs pinned ATT&CK v19.2 matrix"] F --> G["LLM-as-judge review
+ golden gates"] G --> H["Human arbitration
conflict-group corrections"] H --> I[("claude_reviewed_v2
32,100 reviewed rows")] I --> J["Deterministic subset
random.Random(42) → 15,000"] J --> K[("Published
CVE-TO-MITRE V2")] ``` ### Validation applied to the labels | Layer | What it checks | | --- | --- | | Description filter | Vague, short and pre-2005 descriptions rejected before any labelling spend (1,712 dropped) | | Rule-based validator | Technique-id format, per-technique confidence threshold (sub-threshold techniques dropped individually), evidence requirements, existence checks against the pinned ATT&CK v19.2 matrix | | Rule-based auto-correction | Deterministic sanitisation of mapper output (chain re-ordering, generic-technique demotion) before validation | | LLM-as-judge | Independent model verdicts on mapping quality; judge-approved rate tracked per run | | Human arbitration | Conflict-group corrections adjudicated manually; arbitration posture calibrated on a 30-group pilot then generalised with a third-model arbiter | ### Model lineage (what this dataset trained) A 15,000-row superset of this sampling frame (the full reviewed train split) trained a series of encoder classifiers benchmarked under a multi-seed, min-per-seed protocol. The winning configuration — ettin-1B encoder, LoRA rank 32, learning rate 2e-4 — reached **micro F1 0.7607 / exact match 0.4433 on a 4,051-row held-out test set** that is *not* included in this release (kept private, zero description overlap with the public train split). The public dataset therefore has a validated, reproducible downstream use. ## Labels 697 MITRE ATT&CK Enterprise techniques (v19.2) - parent techniques and sub-techniques - constitute the label space. Labelling method, stated honestly: labels were produced by an LLM mapping pipeline (GLM model) with rule-based validation (technique-id format, confidence threshold, evidence requirements, existence checks against the pinned v19.2 matrix), followed by PARTIAL human review of conflict groups; they are model-derived and human-reviewed, not human-authored from scratch. Format: `techniques` lists ids in source order, `technique_names` the matching readable names, and `labels` is a 697-way multi-hot vector whose order is the internal stable label file (a sorted list of technique ids). ## Splits Only the `train` split is published (15,000 rows). The remainder of the reviewed data — 17,100 additional train rows plus the validation (4,045 rows) and test (4,051 rows) splits — is deliberately retained by the author: this dataset represents a substantial amount of work, and the rest is kept for the author's own projects. As a side property, the retained splits have zero description overlap with the published train split (verified by content hashing). ## Sampling Uniform random sampling without replacement: 15,000 rows drawn from the 32,100-row reviewed train split with Python CPython 3.11.9 `random.Random(42).sample` (Mersenne Twister), selected indices emitted in ascending original order. The published train contains no duplicated description. The 15,000-row size is a deliberate release choice by the author (see Splits), not a data-availability constraint. ## Columns | Column | Type | Semantics | | --- | --- | --- | | text | string | CVE description (English, redaction-filtered) | | cve_id | string | CVE id(s); the 166 multi-record rows join their distinct ids with the pipe character | | techniques | list[string] | ATT&CK technique ids in source order | | technique_names | list[string] | readable names aligned with `techniques` | | labels | list[int64] | 697-way multi-hot vector, order = internal stable label file | | attack_version | string | ATT&CK Enterprise version the labels were validated against (19.2) | | epss_score | float64 (nullable) | EPSS score; null when unavailable (max across records on multi-record rows) | | kev | bool | true when any record CVE id is in the CISA KEV catalog as of 2026-10-16 | A canonical JSONL copy of the same rows ships at `data/train.jsonl` for provenance; the parquet file above is the viewer-facing rendering (nulls are genuine: non-finite EPSS values were coerced to null). ## Limitations and Biases - Descriptions are English only; other languages are out of scope. - Long tail of rare labels: 70 of 697 labels have zero support in this subset (full-train support > 0): T1003.007, T1010, T1014, T1016, T1021, T1027.004, T1027.010, T1036.001, T1036.003, T1036.007, T1049, T1052.001, T1059.012, T1069.002, T1071.001, T1071.002, T1098.001, T1098.003, T1114.001, T1119, T1134, T1136.003, T1140, T1176, T1201, T1202, T1218, T1218.005, T1218.015, T1219, T1221, T1491.001, T1496, T1496.001, T1505.001, T1543.002, T1546.002, T1546.008, T1547, T1547.009, T1548, T1548.004, T1548.005, T1557.003, T1560, T1561.001, T1563.002, T1564.010, T1564.012, T1569, T1572, T1574.011, T1578.005, T1580, T1583.002, T1583.004, T1584.001, T1584.004, T1588.002, T1590.001, T1591, T1591.001, T1598.003, T1602, T1608.001, T1619, T1684, T1685.001, T1685.003, T1688. - Labels are model-derived with rule-based validation and partial human review; they carry the biases of that pipeline and should not be treated as ground-truth annotations. - EPSS is partial by design: 18.6% of rows carry a score (capture 2026-09-29 to 2026-09-30); the absence of a score is NOT evidence of a low exploit probability. - KEV membership is a snapshot taken at 2026-10-16T11:33:28+00:00; catalog membership changes over time. - CVSS scores and CWE ids are not included in this v1. ## Citation If you use this dataset, please cite it and the sources it derives from: ```bibtex @misc{cve2attack_v2_2026, author = {SpongeBOB9684}, title = {CVE-TO-MITRE V2: CVE descriptions mapped to MITRE ATT\&CK attack chains}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/SpongeBOB9684/CVE-TO-MITRE-V2}}, license = {CC-BY-4.0}, note = {Labels produced by an LLM mapping pipeline with rule-based validation, LLM-as-judge review and partial human arbitration} } ``` Source attributions: - CVE data: The CVE Program (cve.org), CVE records are distributed free of charge. - Label taxonomy: MITRE ATT&CK Enterprise v19.2, (c) 2026 The MITRE Corporation. This work uses MITRE ATT&CK; MITRE is not affiliated with, nor does it endorse, this dataset.