README.md

DPTrainer

Differential privacy training utilities for PyTorch and Hugging Face Transformers, powered by Opacus.

Overview

DPTrainer provides DPTrainer — a drop-in replacement for Hugging Face's Trainer that adds differential privacy (DP) guarantees via DP-SGD. It handles per-sample gradient clipping, noise injection, privacy budget accounting, and automatic noise calibration so you can fine-tune language models with formal privacy guarantees.

Key Features

• Drop-in Hugging Face integration — DPTrainer extends transformers.Trainer, so all standard training arguments, callbacks, checkpointing, and evaluation workflows work out of the box.
• Automatic noise calibration — specify a target privacy budget (ε, δ) and the noise multiplier is computed automatically.
• Privacy accounting via Opacus — uses Rényi DP (rdp) accounting for end-to-end privacy tracking.
• Gradient clipping strategies — flat, adaptive (AdaClip), and per-layer clipping modes.
• Poisson sampling — optional Poisson sub-sampling for stronger privacy amplification.
• Ghost clipping — memory-efficient per-sample gradient computation via Opacus ghost clipping mode, with built-in safety guards that warn about incompatible method overrides and validate correct loss wrapping at runtime.
• Privacy budget early stopping — training automatically stops when the privacy budget (ε) is exhausted.
• Checkpoint-aware accounting — privacy accountant state is saved and restored with checkpoints for correct budget tracking across restarts.
• privatize_trainer utility — patch any Trainer subclass (e.g., DPOTrainer, Seq2SeqTrainer, SFTTrainer) to use differential privacy without modifying its source code. Note: cannot be used on Trainer itself — use DPTrainer directly for that case.
• Single-GPU training — designed for single-GPU training; distributed training (multi-GPU / multi-node) is not supported.
• Patched components — includes a checkpoint-aware EarlyStoppingCallback compatible with DP training.

Installation

uv pip install --index-url https://europe-west4-python.pkg.dev/jetbrains-ml4se-fed/jbr-fed-python/simple DPTrainer

Or with pip:

pip install --index-url https://europe-west4-python.pkg.dev/jetbrains-ml4se-fed/jbr-fed-python/simple DPTrainer

What's Next?

• Getting Started — quick-start guide with code examples.
• Configuration — full reference for PrivacyArguments and noise calibration.
• Examples — end-to-end training scripts.