sample-training-data/README.md

# Sample training documents

This directory is the authoring area for PiiScan training and retraining examples. It contains one
JSONL file for every required v1 canonical entity type in `plan.md`.

The committed rows are **synthetic format examples only**. They demonstrate the schema and are not a
sufficient training set or a gold benchmark. Do not add real PII or PHI to this repository.

## File organization

- A file is named after its primary canonical type, for example `email_address.jsonl`.
- Every non-empty line is one JSON object and one independently processable document.
- Each file should contain positive examples for its primary type and realistic hard negatives.
- If a document contains other PII, annotate every entity even when it belongs to another file.
- Put multilingual examples in the same type file and identify them with BCP 47 `language` and ISO
  3166-1 alpha-2 `country` values.
- Keep final validation and blind test records outside this folder under access control.

## Required row shape

```json
{
  "id": "globally-unique-record-id",
  "taxonomy_version": "1.0.0",
  "text": "alex.morgan@example.test is the account email.",
  "language": "en",
  "country": "US",
  "domain": "support",
  "entities": [
    {
      "start": 0,
      "end": 24,
      "type": "EMAIL_ADDRESS",
      "subtype": "work"
    }
  ],
  "source": "synthetic-sample",
  "synthetic": true,
  "review_status": "unreviewed",
  "split": "train"
}
```

`subtype` may be `null`. Extra provenance fields are allowed, including `source_revision`,
`template_id`, `subject_group`, `license_id`, `annotator_ids`, and `adjudication_status`.

## Span rules

- `start` is zero-based and inclusive.
- `end` is zero-based and exclusive.
- Offsets address the exact, unmodified value in `text`.
- `text[start:end]` must equal the entity text intended for training.
- Annotate repeated values separately.
- Do not include labels such as `email:` or surrounding punctuation in a value span.
- Annotate the smallest complete semantic span, subject to the taxonomy guidance in `plan.md`.

## Positive and negative examples

A positive row has at least one entity. A hard-negative row has an empty `entities` array and should
look plausibly confusable with the target type. Aim for 30-40% hard negatives in generated training
data, but control the final ratio during dataset assembly rather than duplicating rows here.

Several committed identifiers intentionally use reserved, illustrative, or checksum-invalid values.
They are suitable for demonstrating contextual annotation but not for testing a checksum validator.
Validator-positive suites must use standards-approved test values or controlled ephemeral generation
and must never introduce a real person's identifier.

## Adding data safely

1. Create synthetic or explicitly approved source text.
2. Assign a globally unique, non-identifying `id`.
3. Annotate all spans against the original text.
4. Set accurate provenance and review status. Never mark generated examples as human gold.
5. Keep one generated identity and template family in only one train/validation/test split.
6. Run the validator before dataset assembly:

```powershell
python sample-training-docs/validate.py
```

The validator checks the manifest, file/type mapping, JSON syntax, unique IDs, required fields, and
the original-text span invariant. The future `scripts/build_dataset.py` will perform deeper license,
near-duplicate, split-leakage, and minimum-coverage checks.

## Retraining contract

Training does not consume this directory directly. Every training or retraining run must first build
and validate the canonical dataset:

```powershell
python sample-training-data/validate.py
python scripts/build_training_data.py
python scripts/validate_training_data.py
```

The builder copies valid authored rows without changing text, entities, offsets, or splits; it adds
only missing source-revision provenance. It then normalizes enabled downloaded sources through
`configs/training-data-mappings.json`. The independent validator rejects unknown entity types,
invalid spans, duplicate IDs, cross-split exact-text leakage, stale mapping/schema fingerprints, and
manifest count differences. Model training must read only `prepared-training-data`.