Tutorials

These Jupyter notebooks provide hands-on practice with Opake. Each is self-contained: you build something concrete and see it work.

For conceptual explanations, see the User Guide. For complete function signatures, see the API Reference.

DP-SGD track

Independent per-step Gaussian noise. The classical DP training pipeline.

Tutorial	Task
DP-SGD Training	Train a model with DP-SGD from scratch — the canonical reference.
Privacy Accounting & Calibration	Compose mechanisms, calibrate the noise multiplier, query ε at the end.
Sampling & Microbatching	Plain vs truncated Poisson, microbatch size, the privacy/memory trade-off.
Fine-tuning an LLM	LoRA + adaptive clipping + truncated Poisson on a real Hugging Face model.

DP-FTRL track

Correlated noise across the run via matrix factorization. Better privacy/utility on cumulative updates, at the cost of fixing the training length up front.

Tutorial	Task
DP-FTRL Training	Train a model with DP-FTRL — strategy choice, whole-process accounting, MF noise.

The DP-SGD-track accounting and sampling notebooks share their core ideas with DP-FTRL; the strategy-and-n_steps-up-front difference is covered explicitly in the DP-FTRL training tutorial above.

Cross-cutting

Concerns that apply to either stack.

Tutorial	Task
Privacy Auditing	Validate privacy guarantees empirically with one-run / coin-flip auditing.
Distributed Training	Multi-process DP training on CPU or GPU — synchronized noise, sharded sampling.

Running the tutorials

Install Opake and its tutorial dependencies from a repository checkout:

git clone https://github.com/JetBrains-Research/opake.git
cd opake
uv sync --group examples --all-packages --extra all
jupyter lab docs/tutorials/

The examples group ships Jupyter, matplotlib, torchvision, torchopt, datasets, and wandb; HuggingFace integration comes via --extra all (or --extra transformers). The distributed-training tutorial requires multiple processes via torchrun.