Dataset card: CVSS v3/v4 fix, group split, fields

README.md

@@ -60,13 +60,75 @@ dataset_info:
 
 # Vulnerability-FSTEC
 
-Vulnerability descriptions and severity labels from the [FSTEC](https://vulnerability.circl.lu/recent#fstec), extracted via [Vulnerability-Lookup](https://vulnerability.circl.lu).
+Russian vulnerability descriptions with CVSS scores from the Data Bank of Security Threats (BDU) of [FSTEC of Russia](https://bdu.fstec.ru), extracted via [Vulnerability-Lookup](https://vulnerability.circl.lu/recent#fstec).
+It is used to train the VLAI severity classifier for Russian.
+
+> **Revision a104085 (2026-11-03) fixes the CVSS scores.** Earlier revisions had CVSS 2.0 scores only: BDU stores its v3 and v4 vectors without the `CVSS:3.x/` / `CVSS:4.0/` prefix, and the parser dropped them silently.
+> Reading them changes the severity label of 21,457 of the 82,960 rows of the previous revision (25.9%).
+> This revision also has a group-aware train/validation/test split. Reported and fixed by Christian Bucher, see [VulnTrain #28](https://github.com/vulnerability-lookup/VulnTrain/pull/28).
+
+## Splits
+
+| Split | Rows | Low | Medium | High | Critical |
+|---|---:|---:|---:|---:|---:|
+| train | 78,798 | 2,513 | 31,196 | 32,758 | 12,329 |
+| validation | 10,027 | 288 | 4,209 | 4,154 | 1,376 |
+| test | 9,698 | 290 | 3,928 | 4,033 | 1,447 |
+| **total** | **98,523** | 3,091 (3.1%) | 39,333 (39.9%) | 40,945 (41.6%) | 15,152 (15.4%) |
+
+Two train rows have no usable score and no label.
+
+Rows that share a normalised description or a CVE id are kept in the same split, and each group's split comes from a seeded hash, so a record stays in its split when the dataset is rebuilt.
+The published revision has no description or CVE id shared between splits (checked with [`check_split_leakage.py`](https://github.com/vulnerability-lookup/VulnTrain/tree/main/tools/dataset)).
+
+## Fields
+
+| Field | Description |
+|---|---|
+| `id` | BDU identifier, e.g. `BDU:2021-05687` |
+| `title` | Vulnerability name (Russian) |
+| `description` | Vulnerability description (Russian) |
+| `cvss_v4_0` | CVSS 4.0 base score, base metrics only (CVSS-B), as for the other VLAI datasets (5,037 rows) |
+| `cvss_v3_0` | CVSS 3.x base score (81,177 rows) |
+| `cvss_v2_0` | CVSS 2.0 base score (98,519 rows) |
+| `cvss_v3_1` | Always empty: BDU gives one v3 vector, stored in `cvss_v3_0` |
+| `cve_id` | First CVE id among the record's identifiers in other databases (95,504 rows, 96.9%) |
+| `other_identifiers` | That field as published by BDU, e.g. `CVE-2009-3585, DSA-1944, ID:1400490` |
+| `source` | Always `fstec` |
+| `cpes`, `aliases` | Always empty; kept for a schema shared with the other VLAI datasets |
+
+## Severity labels
+
+The dataset stores scores, not labels. The VLAI trainers take the first available score in the order v4.0, v3.x, v2.0 and map it to Low (< 4.0), Medium (4.0–6.9), High (7.0–8.9) or Critical (≥ 9.0).
+In this revision, 77.3% of the labels come from v3, 17.6% from v2 and 5.1% from v4.
+
+How the vectors are read:
+- BDU's v3 and v4 vectors have no `CVSS:3.x/` / `CVSS:4.0/` prefix; it is added before scoring.
+- v4.0 is scored on its base metrics only (CVSS-B). Threat metrics such as Exploit Maturity are left out.
+- A vector with no impact at all (`C:N/I:N/A:N`, score 0.0) is a placeholder in BDU and is treated as missing.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("CIRCL/Vulnerability-FSTEC", revision="a104085")
+print(dataset)
+```
 
 ## Source
 
-- **Data source**: [Vulnerability-Lookup](https://vulnerability.circl.lu) API
-- **Extraction tool**: [VulnTrain](https://github.com/vulnerability-lookup/VulnTrain)
+- **Data source**: BDU of FSTEC of Russia, via [Vulnerability-Lookup](https://vulnerability.circl.lu)
+- **Extraction tool**: [VulnTrain](https://github.com/vulnerability-lookup/VulnTrain) (`vulntrain-dataset-generation --sources fstec --validation-size 0.1`)
 
 ## Related models
 
 - [CIRCL/vulnerability-severity-classification-russian-ruRoberta-large](https://huggingface.co/CIRCL/vulnerability-severity-classification-russian-ruRoberta-large) — severity classifier trained on this dataset
+
+## Acknowledgements
+
+Thanks to Christian Bucher ([christian140903-sudo](https://github.com/christian140903-sudo)) for finding and fixing the CVSS reading and the `cve_id` field.
+
+## Funding
+
+This work is part of [AIPITCH](https://www.science.nask.pl/en/research-areas/projects/12456).