SAHA-AL: PII Anonymization Benchmark

SAHA-AL is a benchmark for training and evaluating text anonymization systems. It goes beyond detection accuracy by evaluating anonymization as a system under attack — measuring adversarial re-identification risk, contextual privacy leakage, and a formalized privacy-utility tradeoff.

Key Features

• 3 evaluation tasks: PII detection, text anonymization quality, and adversarial privacy risk
• 11 metrics spanning leakage, utility, format preservation, and multi-vector attack resistance
• 8 baseline systems evaluated: 3 rule-based (Regex, spaCy, Presidio) and 5 fine-tuned seq2seq models
• Privacy-Utility Score (PUS): A parameterized framework for navigating the privacy-utility tradeoff
• Failure taxonomy: Diagnostic error classification (full leak, boundary error, format break, context retention)

Dataset

Split	Records	Entities	Avg Entities/Record	PII Density
Train	113,133	286,542	2.53	14.9%
Validation	3,600	9,211	2.56	15.0%
Test	3,600	9,271	2.58	15.2%

Entity types (12): FULLNAME, UNKNOWN, PHONE, EMAIL, ADDRESS, DATE, FIRST_NAME, ID_NUMBER, ZIPCODE, CREDIT_CARD, SSN, ORGANIZATION

Test set is dominated by FULLNAME (57.5%) and UNKNOWN (19.6%).

Data Format

Each record is a JSON object:

{
  "id": "sample_00123",
  "original_text": "Please contact John Smith at john.smith@email.com for details.",
  "anonymized_text": "Please contact Michael Jones at m.jones@mail.org for details.",
  "entities": [
    {"text": "John Smith", "type": "FULLNAME", "start": 15, "end": 25},
    {"text": "john.smith@email.com", "type": "EMAIL", "start": 29, "end": 49}
  ]
}

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("huggingbahl21/saha-al")

# Access splits
train = dataset["train"]       # 113,133 records
val = dataset["validation"]    # 3,600 records
test = dataset["test"]         # 3,600 records (frozen)

# Example record
record = test[0]
print(record["original_text"])
print(record["entities"])

Tasks

Task 1: PII Detection

Input: original_text | Output: detected_entities: [{start, end, type}]

Evaluate span-level precision, recall, and F1 in three modes: exact match, partial (IoU > 0.5), and type-aware.

Task 2: Text Anonymization

Input: original_text | Output: anonymized_text

Replace all PII with realistic synthetic alternatives while preserving non-PII text and document structure.

Task 3: Privacy Risk Assessment

Evaluator-computed over Task 2 outputs. Simulates adversarial attacks against anonymized text to measure residual re-identification risk.

Metrics

Anonymization Quality (Task 2)

Metric	What it measures	Direction
ELR	Entity Leakage Rate — fraction of entities found verbatim in output	↓ lower is better
Token Recall	Fraction of entity tokens absent from output	↑ higher is better
OMR	Over-Masking Rate — non-entity tokens unnecessarily altered	↓ lower is better
FPR	Format Preservation Rate — structured replacements match expected format	↑ higher is better
BERTScore F1	Semantic similarity between original and anonymized text	↑ higher is better

Privacy Risk (Task 3)

Metric	Attack model	Direction
CRR-3	Capitalized 3-gram survival rate (statistical)	↓ lower is better
ERA@k	Entity Recovery Attack — retrieval adversary with candidate pool (SBERT)	↓ lower is better
LRR	LLM Re-identification Rate — generative adversary guesses original entities	↓ lower is better
UAC	Unique Attribute Combination rate — k-anonymity proxy	↓ lower is better

Privacy-Utility Tradeoff

PUS(lambda) = lambda * (1 - ELR/100) + (1 - lambda) * (BERTScore/100)

A single parameterized score where lambda controls the privacy-utility balance (lambda=0.5 gives equal weight).

Leaderboard

Evaluated on the frozen test split (3,600 records, 9,271 entities).

Task 2: Anonymization Quality

System	ELR ↓	Token Recall ↑	BERTScore F1 ↑	PUS (lambda=0.5)
BART-base + PII	0.93%	95.50%	92.74	0.959
Flan-T5-small + PII	0.99%	95.33%	92.47	0.957
T5-small + PII	1.54%	94.15%	92.59	0.955
T5-eff-tiny + PII	4.14%	90.42%	92.57	0.942
DistilBART + PII	1.23%	94.41%	86.34	0.926
spaCy+Faker	26.44%	74.59%	91.86	0.827
Presidio	33.77%	68.18%	90.04	0.781
Regex+Faker	83.39%	23.05%	98.15	0.574

Task 3: Privacy Risk

System	CRR-3 ↓	ERA@1 ↓	ERA@5 ↓	LRR Exact ↓	UAC ↓
BART-base + PII	34.62%	1.90%	4.84%	0.13%	0.33%
Flan-T5-small + PII	34.94%	1.67%	5.08%	—	0.22%
spaCy+Faker	40.62%	9.36%	14.27%	—	1.58%
Presidio	50.33%	20.46%	26.09%	2.12%	1.78%

Task 1: PII Detection

System	Exact F1	Partial F1	Type-aware F1
spaCy+Faker	56.30	69.38	18.18
Presidio	48.57	60.11	4.97
Regex+Faker	25.86	27.99	24.39

Key Findings

1. **Seq2seq models achieve <1% entity leakage** while preserving >92% semantic similarity, dramatically outperforming rule-based systems (26–83% leakage).
2. Structured PII (email, phone, SSN, credit card) is fully solved — 0% leakage across all seq2seq models. Names remain the open challenge, accounting for 83% of residual leakage.
3. Retrieval attacks are more effective than LLM attacks. ERA@1 recovers 1.9% of entities vs LRR's 0.13%, indicating replacement-based anonymization is more vulnerable to database-level adversaries than generative inference.
4. BART-base dominates the Pareto frontier with PUS=0.959, achieving near-maximum privacy (99.1%) and high utility (92.7%).

Evaluation Quick Start

pip install faker spacy bert_score sentence-transformers transformers
python -m spacy download en_core_web_lg

Run a baseline

# Rule-based (regex / spacy / presidio)
python -m baselines.regex_faker_baseline --gold data/test.jsonl --mode spacy --save-spans

# Seq2seq (downloads checkpoint from HuggingFace)
python -m baselines.seq2seq_inference --gold data/test.jsonl --model-name bart-base-pii --output predictions/predictions_bart-base-pii.jsonl

Evaluate

# Task 2: Anonymization quality
python -m eval.eval_anonymization \
    --gold data/test.jsonl \
    --pred predictions/predictions_bart-base-pii.jsonl \
    --print-types

# Task 1: Detection (requires --save-spans from baseline)
python -m eval.eval_detection \
    --gold data/test.jsonl \
    --pred predictions/spacy_spans.jsonl

# Task 3: Privacy risk
python -m eval.eval_privacy \
    --gold data/test.jsonl \
    --pred predictions/predictions_bart-base-pii.jsonl \
    --train data/train.jsonl \
    --skip-lrr

# Bootstrap confidence intervals
python -m eval.bootstrap \
    --gold data/test.jsonl \
    --pred predictions/predictions_bart-base-pii.jsonl \
    --metrics elr token_recall crr3

# Failure taxonomy
python -m analysis.failure_taxonomy \
    --gold data/test.jsonl \
    --pred predictions/predictions_bart-base-pii.jsonl

# Pareto frontier analysis
python -m analysis.pareto_frontier \
    --results Results/all_eval_results.json \
    --plot figures/pareto_frontier.png

# Run the full pipeline
bash scripts/run_all.sh

Submission Format

Task 2 (one JSONL line per record):

{"id": "sample_00000", "anonymized_text": "Please contact Michael Jones at m.jones@mail.org."}

Task 1 (one JSONL line per record):

{"id": "sample_00000", "detected_entities": [{"start": 15, "end": 29, "type": "FULLNAME"}]}

Repository Structure

benchmark/
├── eval/                          # Evaluation modules
│   ├── utils.py                   #   Shared: span matching, format regexes, I/O
│   ├── eval_detection.py          #   Task 1: span P/R/F1
│   ├── eval_anonymization.py      #   Task 2: ELR, Token Recall, OMR, FPR, BERTScore
│   ├── eval_privacy.py            #   Task 3: CRR-3, ERA, LRR, UAC
│   └── bootstrap.py               #   95% bootstrap confidence intervals
│
├── baselines/                     # Baseline systems
│   ├── regex_faker_baseline.py    #   Regex / spaCy / Presidio pipelines
│   ├── seq2seq_inference.py       #   BART / T5 / Flan-T5 / DistilBART inference
│   ├── bert_ner_baseline.py       #   BERT-base token classifier (NER)
│   ├── llm_baseline.py            #   GPT-4o-mini / local LLM zero-shot
│   └── hybrid_baseline.py         #   spaCy detect + LLM replace
│
├── analysis/                      # Analysis & visualization
│   ├── dataset_stats.py           #   Dataset statistics
│   ├── pareto_frontier.py         #   Privacy-utility Pareto + PUS sweep
│   ├── failure_taxonomy.py        #   5-category error classification
│   └── plot_results.py            #   Publication figures
│
├── scripts/                       # Utilities
│   ├── prepare_dataset.py         #   Build train/val/test from raw data
│   ├── push_to_hf.py             #   Push to HuggingFace Hub
│   └── run_all.sh                 #   Full evaluation pipeline
│
├── data/                          # Dataset splits (JSONL)
├── predictions/                   # Model outputs
├── Results/                       # Evaluation JSONs
├── figures/                       # Plots
│
├── RESULTS.md                     # Key results for reporting
├── RESULTS_FULL.md                # Exhaustive results reference
├── METHODOLOGY.md                 # Full methodology & theory guide
└── RESULTS_INTERPRETATION.md      # In-depth results discussion

Trained Model Checkpoints

All seq2seq checkpoints are hosted on HuggingFace:

Repository: [JALAPENO11/pii_identification_and_anonymisations](https://huggingface.co/JALAPENO11/pii_identification_and_anonymisations)

Model	Base Architecture	Parameters	Checkpoint Path
BART-base + PII	facebook/bart-base	139M	checkpoints_pii_aware_loss/bart-base/best_model.pt
Flan-T5-small + PII	google/flan-t5-small	77M	checkpoints_pii_aware_loss/flan-t5-small/best_model.pt
T5-small + PII	t5-small	60M	checkpoints_pii_aware_loss/t5-small/best_model.pt
DistilBART + PII	sshleifer/distilbart-cnn-6-6	230M	checkpoints_pii_aware_loss/distilbart/best_model.pt
T5-eff-tiny + PII	google/t5-efficient-tiny	16M	checkpoints_pii_aware_loss/t5-efficient-tiny/best_model.pt

Benchmark Protocol

1. No test set contamination — do not train or tune on test.jsonl
2. Use the provided evaluation scripts for fair comparison
3. Report compute — GPU type, training time, inference time
4. Disclose pre-training data for LLM-based systems

Citation

If you use SAHA-AL in your research, please cite:

@misc{saha-al-2026,
  title={SAHA-AL: A Multi-Task Benchmark for PII Anonymization with Adversarial Privacy Evaluation},
  author={Mr. A},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/huggingbahl21/saha-al}},
}

Limitations

• Synthetic data: Source texts are generated via Faker-based templates. Results may not fully generalize to real-world PII distributions.
• English only: All data, models, and evaluation metrics are English-specific.
• Entity type ambiguity: ~19.6% of test entities are typed as UNKNOWN due to heuristic type inference limitations.
• Train/test entity overlap: 43.9% of test entity strings appear in training data, which may inflate seq2seq model performance.
• IAA: Inter-annotator agreement was entity-level F1=0.83.

License

This benchmark is released under the MIT License.