license: mit
language:
- en
pretty_name: 'PII-TRACE: Multi-Turn PII NER Dataset'
task_categories:
- token-classification
size_categories:
- n<1K
tags:
- pii
- named-entity-recognition
- span-detection
- conversational
- multi-turn
configs:
- config_name: conversations
default: true
data_files:
- split: train
path: data/train.parquet
PII-TRACE
PII-TRACE is a synthetic dataset for privacy-focused named entity recognition (NER) of personally identifiable information (PII), released as a 500-conversation subset of multi-turn dialogues with exact span annotations.
Dataset summary
This release contains 500 conversations with 2,653 annotated PII spans across nine labels. Among them, 450 conversations contain PII spans and 50 contain none.
Data format
Each record contains:
| Field | Type | Description |
|---|---|---|
id |
string | Unique record ID. |
turns |
list | Ordered turn objects containing turn, user, and assistant. |
spans |
list | Gold PII spans. An empty list indicates an empty control. |
Each item in spans contains:
| Field | Type | Description |
|---|---|---|
label |
string | PII label. |
turn |
integer | Zero-based index of the containing turn. |
source |
string | Message role containing the span (user or assistant). |
start |
integer | Inclusive Unicode code-point start offset in the selected message. |
end |
integer | Exclusive Unicode code-point end offset in the selected message. |
text |
string | Exact annotated substring. |
Do not trim or normalize message text before applying offsets.
Labels
| Label | Description |
|---|---|
private_person |
Person name |
private_date |
Date |
private_url |
Link or route |
private_address |
Mailing or location address |
account_number |
Account or reference number |
private_email |
Email address |
private_phone |
Telephone number |
other_pii |
Other identifier |
secret |
Access code or credential-like value |
Evaluation
Supply turns to the model and keep spans hidden. Report micro and macro exact typed-span precision, recall, and F1 over (turn, source, start, end, label). Untyped exact-span F1 over (turn, source, start, end) may also be reported.
This subset may be used for training, evaluation, and error analysis in PII span detection research.
Limitations
This is a controlled benchmark rather than an estimate of real-world PII prevalence or deployment-level performance. Its English conversations and identifier formats are more regular than production traffic, and current gold spans occur in user messages. The secret label covers credential-like sensitive values but does not assert that every such value is statutory PII.
License
MIT.
Citation
If you use or reference this work, please cite:
@article{zhang2026piitrace,
title = {{PII-TRACE}: A Benchmark for Context-Aware {PII} Detection in Multi-Turn {LLM} Conversations},
author = {Zhang, Kaiyuan and Wang, Chuan and Zhong, Joey and Fryzel, Paul and Polley, Kyle and Ma, Jerry and Li, Ninghui},
journal = {arXiv preprint arXiv:2609.22200},
year = {2026},
url = {https://arxiv.org/abs/2609.22200}
}