---
license: other
license_name: mixed-permissive
license_details: >-
  Same terms as josh-hud/code-pii: each record keeps the license of its base file (MIT,
  Apache-2.0, BSD-2/3-Clause, ISC, Unlicense or CC0-1.0) as redistributed by
  codeparrot/github-code-clean. Generator code Apache-2.0.
task_categories:
  - token-classification
language:
  - code
  - en
tags:
  - secrets-detection
  - credential-detection
  - pii
  - code
  - synthetic
  - span-labels
  - coreference
  - anonymization
  - preview
pretty_name: code-pii preview (36% cluster-preserving sample of code-pii v3)
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet
---

# code-pii-preview

A seeded **36% cluster-preserving sample** of [josh-hud/code-pii](https://huggingface.co/datasets/josh-hud/code-pii) v3: 4.28 GB of zstd Parquet instead of 11.92 GB. It has the same schema, labels, entity tables and license terms as the parent.

## Sampling

Every complete cluster is retained iff `blake2b("42:" + cluster_id) / 2^64 < 0.36`. The decision is reproducible and order-independent. It never splits a cross-document entity cluster; all files and the embedded entity table for a retained cluster remain together.

## Contents

| path | description |
|---|---|
| `data/train-AAAAA-BBBBB.parquet` | 8 zstd Parquet files; clusters are contiguous and never cross files |
| `sample.jsonl` | approximately 500 records comprising complete clusters, same schema |
| `stats.md` | totals, file list and label mix |

## Numbers

| metric | preview | full v3 | ratio |
|---|---:|---:|---:|
| records | 1,900,308 | 5,278,400 | 0.360 |
| clusters | 856,252 | 2,379,311 | 0.360 |
| files with >=1 span | 1,256,231 (66.1%) | 3,491,739 (66.2%) | 0.360 |
| labeled spans | 2,853,437 | 7,936,370 | 0.360 |
| hard-negative spans | 1,781,642 | 4,950,684 | 0.360 |
| Parquet bytes | 4,284,943,933 | 11,915,030,320 | 0.360 |

Labels: `api_key` 37.1%, `password` 17.9%, `token` 13.1%, `email` 11.8%, `private_key` 8.0%, `person_name` 7.3%, `username` 4.7%.

## Schema

Each record includes the v2 fields plus `cluster_id`, `cluster_kind`, `cluster_size`, `doc_index`, and an embedded `entities` table on `doc_index == 0`. Each positive span adds `entity_id`, `surface_form_type`, and a gold `replacement`. Spans sharing an `entity_id` across files are the same entity and must receive consistent replacements. Full field documentation is in the [parent data dictionary](https://huggingface.co/datasets/josh-hud/code-pii/blob/main/DATA_DICTIONARY.md).

## Loading

```python
from datasets import load_dataset
ds = load_dataset("josh-hud/code-pii-preview", split="train")  # or streaming=True
rec = ds[0]
for s in rec["spans"]:
    assert rec["text"][s["start"]:s["end"]] == s["value"]
```

The parent generator's `validate.py` reports zero offset, entity, replacement, ownership, contiguity, collision or format problems on the regenerated v3 data. Use the parent's `eval_creddata.jsonl` for real-code detection evaluation; use held-out synthetic clusters for coreference and exact anonymization evaluation.

## Caveats

All parent caveats apply: injected-line placement artifacts; unlabeled real secrets, emails and names in the base corpus; unrelated files in `bundle` clusters; literal rather than indirect cross-file recurrence; and a Java/JavaScript/HTML-heavy language mix.