TIDE 2.0

A data de-identification and anonymization toolkit that combines multiple anonymization strategies with cryptographic techniques and machine learning-based entity recognition.

Get Started

Dev Container (recommended)

The repository includes a Dev Container configuration that sets up the full development environment automatically: Python 3.12, uv, all dependencies (including GPU group), pre-commit hooks.

Prerequisites:

• Docker
• VS Code with the Dev Containers extension

Steps:

1. Clone the repository and open it in VS Code:

   git clone https://github.com/susom/tide2-core.git
   cd tide2
   

2. When VS Code detects .devcontainer/devcontainer.json, click Reopen in Container (or run the command Dev Containers: Reopen in Container from the command palette).
3. The virtual environment at /opt/tide2-core/.venv is activated by default in all terminals.

The Dev Container includes these VS Code extensions pre-installed: Python, Ruff, Jupyter, Docker, and TOML support.

Local Installation (without Dev Container)

If you prefer to develop outside the Dev Container:

# Install uv (https://docs.astral.sh/uv/getting-started/installation/)
# macOS and Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and install
git clone https://github.com/susom/tide2.git
cd tide2

uv python install 3.12.8
uv sync

# Activate the virtual environment before running any Python commands
source .venv/bin/activate          # macOS / Linux
# .venv\Scripts\activate           # Windows (PowerShell)

Quick Start: Interactive Tutorial

The tutorial notebook walks you through the de-identification pipeline step by step.

In the Dev Container (or local VS Code):

1. Open notebooks/tide2_pipeline.ipynb (the Jupyter extension is pre-installed in the Dev Container)
2. When prompted for a kernel, select the .venv (Python 3.12) environment

Jupyter in the browser (local installation):

uv sync --group dev              # Install Jupyter (dev dependency group)
source .venv/bin/activate
jupyter notebook notebooks/tide2_pipeline.ipynb

View the notebook on GitHub: TIDE 2.0 Pipeline Tutorial

Troubleshooting:

• Run from the repo root — launch Jupyter from the tide2/ directory so that relative paths resolve correctly.
• GCP credentials are not required — the notebook downloads the transformer model from HuggingFace Hub by default. Set project_id and bucket_name in the Configuration cell only if you want to use GCS-hosted weights.
• Kernel crashes — if the Jupyter kernel crashes repeatedly, restart Jupyter (Ctrl+C, then re-launch) and run the cells from the top.

Visualizer Preview

TIDE 2.0 includes a Streamlit visualizer for comparing original and de-identified text side by side:

Launch it with:

tide2-visualizer

To stop the visualizer, press Ctrl+C in the terminal (works on macOS, Linux, and Windows).

Overview

TIDE 2.0 is a Python package for anonymizing sensitive data in healthcare and research contexts. It identifies and anonymizes personally identifiable information (PII) while maintaining data utility for analysis and research.

Features

Entity Recognition

• Transformer-based NER: HuggingFace transformer models with direct batch inference (bypasses HF pipeline), BIO token aggregation, and chunk-to-document reassembly
• Regex recognizers: Phone, URL/IP, Email, SSN, Address — replacements for Presidio defaults (10-100x faster)
• Healthcare-specific: MRN, Accession Number, HAR code recognizers
• Known values detection: Aho-Corasick based matching against patient databases
• Specialized: Base64 image detection, genetic sequence detection, LLM-based JSON recognizer
• Cached results: Pre-computed NER results from GPU batch processing via CachedResultsTransformerRecognizer
• Presidio Integration: Built on Microsoft's Presidio framework

Anonymization Strategies

• HIPS (Healthcare Identity Protection System): Cryptographic deterministic anonymization for names, locations, and alphanumeric identifiers
• Accession number hashing: SHA256-based, compatible with BigQuery UDF
• Faker Integration: Realistic fake data generation
• Date Jittering: Deterministic, privacy-preserving date shifts derived from patient keys
• Age Grouping: Age range categorization

Cryptographic Protection

• Format-Preserving Encryption (FPE): Maintains data format during encryption
• Key Management: Key generation, storage, and derivation utilities
• Deterministic date jitter: Batch-capable date shift derivation from cryptographic keys
• String Selection: HMAC-based cached string selection

Ray-based Batch Processing

• Runner module: Single-node job runner with local and VM modes via tide2-runner CLI
• Ray actors: RecognizerActor, AnonymizerActor, TransformerInferenceActor, BIOAggregationActor, ReassemblyActor for ray.data.map_batches
• Two-stage GPU/CPU pipeline: GPU inference returns raw BIO tokens; CPU actors aggregate them concurrently via Ray Data streaming
• Direct inference: Bypasses HuggingFace pipeline dispatch loop with batch tokenize → single GPU forward pass → offset-based extraction
• Adaptive GPU batching: Auto-computes batch size from model config and free GPU memory; adjusts based on text lengths with VRAM-aware budgets (override via --short-seq-budget)
• OOM recovery: Automatic batch splitting on CUDA out-of-memory errors
• Fault tolerance: Actor restarts, task retries, graceful shutdown
• YAML config: All CLI arguments can be specified in a YAML config file (--config)

Utilities

• Text processing: Text chunking, BIO aggregation, span reconstruction, deduplication
• String parsers: Name parsing/classification, address parsing, format detection
• Span metrics: Gold vs ML evaluation, O(n log n) conflict resolution
• GCS cache: Auto-download models from GCS to ~/.cache/tide2/
• Model compilation: torch.compile with mega-cache support for faster inference startup

Command Line Tools

• tide2-runner: Ray-based single-node job runner with six job types: recognizer, anonymizer, transformer, reassembly, pipeline (full end-to-end), and llm-recognizer. Supports YAML config files (--config) and dry-run mode (--dry-run).
• tide2-visualizer: Streamlit app for side-by-side PHI comparison and entity editing.

Cloud Integration

• GCS: input/output I/O and model caching.
• BigQuery: input/output of notes and recognizer/anonymizer results (e.g. via ARRAY_AGG-grouped chunk columns) for the runner and visualizer.
• Automatic Caching: Download and cache models from GCS automatically ($TIDE_CACHE_DIR).

CLI Usage

Runner CLI (Ray-based processing)

# Run recognition locally
tide2-runner run recognizer -i ./data/input -o ./data/output

# Run with more resources (e.g. on a large VM), reading/writing from GCS
tide2-runner run recognizer -i gs://bucket/input -o gs://bucket/output \
    --num-cpus 224 --num-actors 200

# Run transformer NER on GPU
tide2-runner run transformer -i ./data/input -o ./data/transformer_output \
    --model StanfordAIMI/stanford-deidentifier-v2 --batch-size 2048

# Run transformer with YAML config
tide2-runner run transformer --config config.yaml

# Run the full pipeline (transformer -> recognizer -> anonymizer)
tide2-runner run pipeline -i ./data/input.parquet -o ./data/output \
    --model StanfordAIMI/stanford-deidentifier-v2

# If you are running on Mac, you can use --object-store-gb option to set
tide2-runner run pipeline -i ./data/input.parquet -o ./data/output \
     --model StanfordAIMI/stanford-deidentifier-v2  --object-store-gb 2

# Run anonymization
tide2-runner run anonymizer -i ./data/recognized -o ./data/anonymized \
    --salt /path/to/salt.bin --key /path/to/key.bin

# Run on a small box (e.g. 2-CPU Google Colab) WITHOUT deadlocking. Two fixes
# are required together (see below): fractional CPUs AND --no-checkpoint.
# GPU box (T4): the transformer actor is GPU-pinned, so budget read/flat-map/
# write/agg fractionally; CPU-only box: also give the transformer actor ~C-1.
tide2-runner run pipeline -i ./data/input.parquet -o ./data/output \
    --model StanfordAIMI/stanford-deidentifier-v2 \
    --num-actors 1 --cpus-per-actor 0.5 --worker-num-cpus 1.0 \
    --read-cpus 0.25 --flat-map-cpus 0.25 --write-cpus 0.25 \
    --agg-num-cpus 0.5 --transformer-cpus 0.25 --no-checkpoint

Why small boxes deadlock (and how to size knobs by hardware)

Ray Data runs every operator of a stage concurrently and, under Ray 2.55's reservation allocator, must reserve a minimum CPU slice for every eligible operator at once. When that sum exceeds the cluster's CPUs, nothing schedules and the stage hangs forever at 0/1 (backpressured:tasks(ResourceBudget)). On a 2-CPU box there are two independent causes — both must be fixed together:

1. Whole-CPU operator reservations. Defaults reserve ~1 CPU per operator; read + flat_map + actor + agg + write exceeds 2. Fix with fractional CPUs.
2. The checkpoint shuffle. Row-level resume injects a sort + repartition shuffle (extra operators) that re-triggers the deadlock even with fractional CPUs. Fix with --no-checkpoint (trades resume capability, not correctness).

The knobs are additive and default to today's whole-CPU reservations + checkpointing on, so omitting them preserves large-VM behavior. Size them to fit the sum of a stage's concurrent operator reservations within the available CPUs (C = total CPUs):

• Big box (C ≳ 16): use defaults (omit all knobs).
• Transformer stage: --read-cpus, --flat-map-cpus, --write-cpus, --agg-num-cpus (BIO aggregation actor), --transformer-cpus (CPU floor for the transformer actor; leave unset on GPU, set to ~C - 1 on CPU-only boxes — it also caps the actor's torch threads).
• Recognizer / anonymizer stages: --cpus-per-actor (supervisor), --worker-num-cpus (worker actor), --read-cpus, --write-cpus. Each pool slot needs supervisor + worker CPUs, so budget both.
• All stages on C ≲ 4: add --no-checkpoint.

Interactive Visualizer

# Launch the Streamlit PHI visualizer
tide2-visualizer

Docker Images

Several targets are built from a single multi-stage Dockerfile:

• production-cpu — slim CPU-only image (no CUDA). Used by recognizer, anonymizer, and BigQuery tasks.
• production-gpu — GPU image based on nvidia/cuda:13.0.2-cudnn-runtime-ubuntu24.04. Used by transformer inference. (The ML stack — torch, transformers, spacy — ships in both images, since it is a required core dependency.)
• development — Dev Container target with git, gcloud, build tools, and the full dev environment.
• test — extends development and runs the test suite (used by make test-docker).

Build and push the GPU image (requires DOCKER_REGISTRY and DOCKER_IMAGE_GPU in .env):

make docker         # build + push the GPU image (alias for docker-gpu)
make docker-gpu     # build + push the GPU image
make test-docker    # build the test target and run the suite in Docker

Dependency Groups

• llm: LLM provider SDKs for the optional LLM-based recognizer (anthropic, openai, google-genai, google-cloud-aiplatform)
• dev: Development tools (pytest, pytest-cov, ty, ruff, pre-commit), Jupyter, and the evaluation libraries (scikit-learn, scipy, tqdm)
• evaluation: Evaluation/analysis libraries (scikit-learn, scipy, tqdm)
• test: Minimal test dependencies (pytest, pytest-cov)
• docs: API documentation generation (pdoc)

Install an optional group as an extra with uv sync --extra <name>, or all extras with uv sync --all-extras. (These same sets are also defined as [dependency-groups], usable with uv sync --group <name>.)

Note: The full ML inference stack (torch, transformers, spacy) ships in the main package by default — it is required, since no model can run without it. The llm extra is only needed for the optional LLM-based recognizer.

Architecture

tide2/
├── recognizers/              # PII detection (Presidio EntityRecognizer subclasses)
├── anonymizers/              # PII replacement (Presidio Operator subclasses)
├── transformers/             # Core NER inference engine (TransformerCore)
│   ├── core.py              # Model loading, direct inference, BIO aggregation
│   └── config.py            # Model configuration management
├── actors/                   # Ray actors for distributed batch processing
│   ├── transformer.py       # GPU inference actor + CPU BIO aggregation actor
│   ├── recognizer.py        # CPU recognizer actor
│   ├── anonymizer.py        # CPU anonymizer actor
│   ├── reassembly.py        # Chunk-to-document reassembly actor
│   └── llm_recognizer.py    # LLM-based recognizer actor
├── cryptographic/            # FPE, key management, date jitter derivation
├── string_parsers/           # Name/address parsing, format detection
├── runner/                   # Ray-based single-node job runner + CLI
│   ├── local_runner.py      # LocalJobRunner: transformer/recognizer/anonymizer/reassembly/pipeline/llm
│   ├── cli.py               # tide2-runner CLI with YAML config support
│   ├── transformer.py       # Document chunking and reassembly logic
│   ├── fault_tolerance.py   # Actor restarts, graceful shutdown
│   └── utils.py             # Runner utilities
├── cli/                      # Streamlit visualizer
├── utils/
│   ├── gcs_resource_manager.py  # GCS auto-download and caching
│   ├── gcs_connector.py        # GCS file I/O
│   ├── span_metrics.py         # Evaluation metrics and conflict resolution
│   ├── text_processing.py      # Chunking, BIO aggregation, span reconstruction
│   ├── serialization.py        # RecognizerResult <-> dict conversions
│   ├── llm_model.py            # LLM client utilities
│   ├── batch_columns.py        # Batch column constants
│   ├── constants.py            # Shared constants
│   └── resource_utils.py       # Resource path helpers
└── resources/                # Config files (model configs, name lists, etc.)

Testing

# Run all unit tests (coverage report prints automatically)
uv run pytest

# Run without coverage (faster, useful when debugging)
uv run pytest --no-cov

# Run a specific test file
uv run pytest tests/test_masking_anonymizer.py

# Skip slow integration tests
uv run pytest -m "not integration"

Coverage is configured in pyproject.toml and runs automatically with pytest. Three reports are generated on each run:

• Terminal: line-by-line missing coverage printed to stdout
• HTML: detailed report at htmlcov/index.html
• XML: coverage.xml (Cobertura format)

Documentation

API Reference

API documentation is hosted via GitHub Pages: https://susom.github.io/tide2-core/

To build or preview docs locally (generated with pdoc):

# Install docs dependencies. pdoc imports every module (including the LLM
# utilities), so the `llm` extra is required in addition to `docs`. The ML
# stack (torch/transformers/spacy) ships in the base install.
uv sync --extra docs --extra llm

# Live preview (opens a local server with hot reload)
make docs-serve

# Generate static HTML to docs/
make docs

Deployment to GitHub Pages is automated: the .github/workflows/docs.yml workflow runs make docs on every push to main and publishes the docs/ directory as a Pages artifact.

Other Resources

• Examples: Check the notebooks/ directory for usage examples
• Tests: Test suite in tests/ directory

Requirements

• Dev Container: Recommended — provides the full environment with no manual setup (requires Docker and VS Code with the Dev Containers extension)
• Python: 3.12 (required, >=3.12,<3.13) — constrained to 3.12 for compatibility with the spacy/thinc C-extension stack and other pinned dependencies.
• Package Manager: uv (not pip or poetry)
• Virtual Environment: .venv/ (activated automatically in the Dev Container; must be activated manually for local installs)
• Core Dependencies: Presidio, Ray (>=2.54), Cryptography, Faker, Google Cloud libraries, and the ML inference stack (torch, transformers>=5.0, spacy) — all required and shipped in the base install

Security Considerations

• Cryptographic operations use standard libraries (cryptography, pyca/cryptography)
• Format-preserving encryption maintains data format during encryption
• Key management supports generation, storage, and rotation
• Anonymization strategies are designed to prevent re-identification

Contributing

Please see CONTRIBUTING.md for the development workflow, branching model, commit-message conventions, and the pull request checklist. The Dev Container setup above provisions the full development environment (including pre-commit hooks) automatically.

License

This project is licensed under the MIT License - see the LICENSE-MIT file for details.

Citation

If you use TIDE 2.0 in your research, please cite:

@software{tide2,
  title={TIDE 2.0: Data De-identification and Anonymization Toolkit},
  author={TIDE 2.0 Team},
  year={2025},
  url={https://github.com/susom/tide2}
}

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Development: See the Contributing section above

Synthetic Data Notice: All sample data included in this repository (under notebooks/sample_data/) is entirely synthetic and fabricated. No real patient data is included. See notebooks/sample_data/README.md for details.

Note: This toolkit is designed for research and development purposes. Please ensure compliance with relevant privacy laws and regulations (HIPAA, GDPR, etc.) when using in production environments.