---
license: cc-by-nc-4.0
language:
  - en
  - de
  - fr
  - es
  - ru
task_categories:
  - text-classification
  - token-classification
tags:
  - security
  - secret-detection
  - credentials
  - dlp
  - code
  - pii
size_categories:
  - 100K<n<1M
configs:
  - config_name: corpus
    data_files: corpus.parquet
    default: true
  - config_name: prowlbench
    data_files: prowlbench.parquet
---

<p align="center"><img src="https://huggingface.co/datasets/Podric/prowl-secrets-corpus/resolve/main/logo.png" width="360" alt="Prowl"></p>

# Prowl secrets corpus

A labeled corpus for training and evaluating secret detectors (credentials, API keys, tokens,
database URIs, private keys, and passwords) across **code, Jira tickets, Confluence pages, chat,
and logs**, in multiple languages. It is the training data behind
[Prowl](https://github.com/Lercas/prowl) and its stage-3
[encoder](https://huggingface.co/Podric/prowl-secret-encoder).

**503,027 records: 181,530 positive, 321,497 negative.**

> No live credentials. Every value is synthetic, format-preserving-obfuscated, or drawn from a public
> test fixture. The corpus is built so it can be handled without leaking a real secret.

## Loading

```python
from datasets import load_dataset

ds = load_dataset("Podric/prowl-secrets-corpus")                    # full corpus (default config)
bench = load_dataset("Podric/prowl-secrets-corpus", "prowlbench")   # the held-out evaluation config

ex = ds["train"][0]   # {"text", "value", "label_binary", "label_type", "span", "source", "origin", "context"}
```

## Fields

| field | type | meaning |
|---|---|---|
| `text` | string | the fragment to classify (code line, config value, ticket/log/chat text) |
| `value` | string \| null | the secret substring, when positive |
| `label_binary` | 0 \| 1 | contains a secret |
| `label_type` | string \| null | secret type (`aws_access_key_id`, `stripe_secret_key`, `generic_password`, ...) |
| `span` | [int,int] \| null | `[start,end)` of `value` inside `text` (for NER) |
| `source` | string | channel: `code` \| `jira` \| `confluence` \| `log` \| `slack` |
| `origin` | string | provenance tag (see below) |
| `context` | string \| null | surrounding context, when available |

## Composition by channel

| code | Jira | Confluence | log | chat |
|:--:|:--:|:--:|:--:|:--:|
| 302,153 | 97,122 | 80,092 | 19,883 | 3,777 |

Non-code (tickets, wiki, logs, chat) is ~40% of the corpus, the realistic-leak surface that
code-only datasets omit.

## Provenance: where each part comes from

| origin | share | source | how it was made |
|---|:--:|---|---|
| `synthetic` | 73.8% | original | format- and checksum-correct fake secrets generated per type, injected into templated carriers at known spans |
| `creddata` | 8.4% | [Samsung/CredData](https://github.com/Samsung/CredData) | real-world credential samples, **character-class obfuscated** so format survives but no real value does |
| `hf` | 6.8% | [ai4privacy/pii-masking-200k](https://huggingface.co/datasets/ai4privacy/pii-masking-200k), [nvidia/Nemotron-PII](https://huggingface.co/datasets/nvidia/Nemotron-PII) | synthetic PII prose; passwords → positives, other PII → hard negatives (`neg_pii`) |
| `fleet*` / `authored` / `carrier` | 10.9% | original | synthetic carriers: code in 17 languages, 44 config formats, ticket/wiki/chat/log documents, hard negatives, obfuscation |
| `fixture` | <0.1% | gitleaks / detect-secrets / whispers test suites (public) | imported public test fixtures |

The `synthetic`, `fleet*`, `authored`, and `carrier` portions are original to this project. The
`creddata`, `hf`, and `fixture` portions are derived from the linked upstream datasets and remain
under their respective licenses (CredData; ai4privacy and Nemotron-PII per their HF cards; the tool
fixtures per each project's license). Attribute the originals if you redistribute.

## Splits and leakage control

The corpus is meant to be split **value-disjoint** (no secret value appears in both train and eval)
and **origin-disjoint**: the real-secret sources (`creddata`, `hf`) are held out so a validation
score measures generalization, not memorization. The encoder's operating threshold is calibrated on
such a held-out split.

## `prowlbench` config

The second config is [ProwlBench](https://github.com/Lercas/prowlbench):
3,843 leakage-safe evaluation cases (structured tokens, generic high-entropy keys, multilingual
free-form prose) with per-case tier and language, used to compare scanners as real subprocesses.
It is value-disjoint from `corpus`.

## Intended use

Training and evaluating secret/credential detectors and DLP classifiers. Not a source of real
credentials and not usable as one.

## License

Noncommercial use only (CC BY-NC 4.0). Not for use in commercial products. The original portions
(`synthetic`, `fleet*`, `authored`, `carrier`) are released under CC BY-NC 4.0. The derived portions
(`creddata`, `hf`, `fixture`) stay under their upstream licenses, listed in the provenance table.
Attribute the originals if you redistribute.

## Citation

```bibtex
@misc{prowl_secrets_corpus,
  title        = {Prowl secrets corpus},
  author       = {Prowl},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/Podric/prowl-secrets-corpus}
}
```