---
license: cc-by-4.0
task_categories:
- text-classification
language:
- en
pretty_name: Vulnerability scores (CVSS v3 combined)
size_categories:
- 100K<n<1M
tags:
- vulnerability
- cybersecurity
- security
- cve
- cvss
dataset_info:
  features:
  - name: id
    dtype: string
  - name: title
    dtype: string
  - name: description
    dtype: string
  - name: cpes
    sequence: string
  - name: cvss_v4_0
    dtype: float64
  - name: cvss_v3_1
    dtype: float64
  - name: cvss_v3_0
    dtype: float64
  - name: cvss_v2_0
    dtype: float64
  - name: patch_commit_url
    dtype: string
  - name: source
    dtype: string
  - name: cvss_v3_1_v3_combined
    dtype: float64
  - name: severity_band
    dtype: string
  splits:
  - name: train
    num_bytes: 144500000
    num_examples: 509067
  - name: test
    num_bytes: 16100000
    num_examples: 56502
  download_size: 160600000
  dataset_size: 160600000
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

# Vulnerability scores (CVSS v3 combined)

A labeled slice of [CIRCL/vulnerability-scores](https://huggingface.co/datasets/CIRCL/vulnerability-scores) for training and evaluating models that predict CVSS v3 severity from a vulnerability description.

Every row has a combined v3 score and a severity band. Rows with no v3.1 or v3.0 score were dropped.

## What changed from the original

The CIRCL dataset stores four separate CVSS columns (`cvss_v4_0`, `cvss_v3_1`, `cvss_v3_0`, `cvss_v2_0`). Those versions are not on the same scale, so this derivative does **not** average them or prefer v4.

`cvss_v3_1_v3_combined` is:

1. `cvss_v3_1` when it is present
2. otherwise `cvss_v3_0`

v3.0 and v3.1 are the same scoring system (mean absolute difference about 0.09 on overlapping rows). v2 and v4 are not mixed in.

`severity_band` is the standard CVSS qualitative rating of that combined score:

| Band | Score |
| --- | --- |
| `low` | 0.0–3.9 |
| `medium` | 4.0–6.9 |
| `high` | 7.0–8.9 |
| `critical` | 9.0–10.0 |

Unlabeled rows are removed. That includes rows with no CVSS score at all, and rows that only have a v2 or v4 score.

Original columns are kept, including the unused v2/v4 values.

Train and test splits follow CIRCL’s split. A row is not moved between splits.

## Splits

| Split | Rows | From v3.1 | From v3.0 fallback |
| --- | ---: | ---: | ---: |
| train | 509,067 | 410,050 | 99,017 |
| test | 56,502 | 45,623 | 10,879 |
| total | 565,569 | 455,673 | 109,896 |

CIRCL’s full dump is 779,178 rows. This set keeps 565,569 (72.5%).

### Severity bands

| Split | low | medium | high | critical |
| --- | ---: | ---: | ---: | ---: |
| train | 18,546 | 222,832 | 206,725 | 60,964 |
| test | 2,041 | 24,767 | 23,017 | 6,677 |

## Fields

Same as CIRCL, plus:

| Field | Type | Description |
| --- | --- | --- |
| `cvss_v3_1_v3_combined` | float | v3.1 score, or v3.0 if v3.1 is missing. Never null in this dataset. |
| `severity_band` | string | `low`, `medium`, `high`, or `critical`. Never null in this dataset. |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("AgileRLArena/vulnerability-scores-cvss-v3")
print(ds["train"][0]["description"])
print(ds["train"][0]["cvss_v3_1_v3_combined"], ds["train"][0]["severity_band"])
```

Typical training target is `severity_band` (classification) or `cvss_v3_1_v3_combined` (regression). Input is `description`.

## Source and license

Derived from [CIRCL/vulnerability-scores](https://huggingface.co/datasets/CIRCL/vulnerability-scores) (CIRCL / Vulnerability-Lookup, [VulnTrain](https://github.com/vulnerability-lookup/VulnTrain)). Paper: [VLAI: A RoBERTa-Based Model for Automated Vulnerability Severity Classification](https://arxiv.org/abs/2507.03607).

License is **CC-BY-4.0**, same as the original. Credit CIRCL when you use this dataset.