🏝️ Privasis

Welcome! This is the official repository for our paper "Privasis: Synthesizing the Largest 'Public' Private Dataset from Scratch". Please visit our website for an overview of the project. We will soon release our dataset and models. Stay tuned!

Environment Setup

conda env create -f environment.yml; conda activate privasis
huggingface-cli login

---

Quick Start

You can use one of the shell scripts for easy launch

./run_generate.sh  # builds Privasis
./run_sanitize.sh  # builds Privasis-Sanitization on top of the generated Privasis. You first need to generate some Privasis data

This section covers the two main ways to run models for the Privasis pipeline:

1. OpenAI API - simplest setup, pay-per-use
2. vLLM Server (local models) - run models on your own GPUs

Option 1: OpenAI API

Set your API key and run directly:

export OPENAI_API_KEY="your-api-key"

# Generate Privasis with GPT-4.1
python generate.py \
    --run-id my-run \
    --n_seeds 100 \
    --generator-model gpt-4.1

# Generate Privasis-Sanitization with GPT-4.1 on the generated Privasis data
python sanitize.py \
    --privasis-data-id my-run \
    --run-id my-run-sanitized \
    --sanitization-model gpt-4.1

Option 2: vLLM Server (Local Models)

For local models, start a vLLM server and use parallel workers for optimal throughput.

Step 1: Start the vLLM server

# Basic usage (1 GPU, port 8000)
./start_vllm_server.sh openai/gpt-oss-120b

# With 4 GPUs on port 8080
./start_vllm_server.sh openai/gpt-oss-120b 4 8080

# With custom memory utilization
./start_vllm_server.sh openai/gpt-oss-120b 1 8000 0.95

Step 2: Run generation with parallel workers

python generate.py \
    --generator-model openai/gpt-oss-120b \
    --vllm-server-url http://localhost:8000/v1 \
    --num-workers 8 \
    --n_seeds 1000 \
    --run-id parallel-generation

Step 3: Run sanitization with parallel workers

python sanitize.py \
    --privasis-data-id parallel-generation \
    --sanitization-model openai/gpt-oss-120b \
    --vllm-server-url http://localhost:8000/v1 \
    --num-workers 8 \
    --run-id parallel-sanitization

vLLM Server Architecture

┌─────────────────────────────────────────────────────────┐
│  vLLM Server (single instance, holds model in GPU)      │
│  http://localhost:8000/v1                               │
└───────────────────────────┬─────────────────────────────┘
                            │ HTTP requests
        ┌───────────────────┼───────────────────┐
        ▼                   ▼                   ▼
   ┌─────────┐         ┌─────────┐         ┌─────────┐
   │ Thread 1│         │ Thread 2│         │ Thread N│
   └─────────┘         └─────────┘         └─────────┘
                            │
                    ┌───────▼───────┐
                    │ Single output │
                    │ file (locked) │
                    └───────────────┘

Tips for vLLM Server

• Number of workers: Start with 4-8 workers and adjust based on GPU utilization
• Memory: Reduce --gpu_memory_utilization if you see OOM errors
• Model name consistency: Use the same model name for both the server and client scripts
• Monitoring: Watch the vLLM server output for request queuing and throughput metrics

---

Supported Providers

Provider	Model Patterns	Environment Variable
OpenAI	gpt-4*, gpt-5*, o1-*, o3*	OPENAI_API_KEY
NVIDIA NIM	nvdev/*	NVDEV_API_KEY
vLLM Server	Any model + --vllm-server-url parameter	N/A

---

Pipeline Overview

The pipeline has two stages:

1. Generate Privasis (generate.py) - Generate synthetic records with PII
2. Sanitize Privasis (sanitize.py) - Abstract or remove target information

1. Build Privasis (generate.py)

Generates synthetic records containing rich privacy-sensitive information

Command-Line Arguments

Argument	Default	Description
--run-id	v0	Run ID (used as output filename under outputs/privasis/)
--n_seeds	10	Number of seeds to generate profiles for
--generator-model	(required)	Model to use for generation
--seeds_path	None	Path to custom seeds file for profile generation
--num_events	1	Number of events per profile
--tolerance	3	Number of retries before accepting failure
--print	False	Print generated outputs to console
--vllm-server-url	None	URL of vLLM server (e.g., http://localhost:8000/v1)
--num-workers	1	Number of parallel worker threads
--embedding-model	Qwen/Qwen3-Embedding-0.6B	Embedding model for diversity scoring (HuggingFace or OpenAI)
--embedding-device	cuda	Device for HF embedding model (cpu, cuda, cuda:0, etc.)

Note: The embedding model is used for computing diversity scores (Vendi score) during Metropolis-Hastings sampling. Supported models include HuggingFace models (e.g., Qwen/Qwen3-Embedding-0.6B, Qwen/Qwen3-Embedding-4B) or OpenAI models (e.g., text-embedding-3-small).

Output

Results are saved to outputs/privasis/{run-id}/:

• {run-id}.jsonl - Generated records with PII

---

2. Build Privasis-Sanitization (sanitize.py)

Sanitizes (abstracts or removes) the target information in the generated records.

Command-Line Arguments

Argument	Default	Description
--privasis-data-id	mark07	Run ID from generation phase
--run-id	v0	Run ID for this sanitization run
--sanitization-model	openai/gpt-oss-120b	Model to use for sanitization
--retry-limit	1	Number of retries if sanitization fails
--attr-selection-weighting	sensitivity	Attribute selection strategy (uniform or sensitivity)
--print	False	Print sanitized outputs to console
--vllm-server-url	None	URL of vLLM server
--num-workers	1	Number of parallel worker threads

Output

Results are saved to outputs/sanitized_privasis/:

• {privasis-data-id}_{run-id}.jsonl - Sanitized records
• {privasis-data-id}_{run-id}_errors.jsonl - Failed instances (if any)

---

Evaluation

Coming soon with our dataset release!

---

BibTeX

Please cite our work if you find this repo useful.

@article{kim2026privasis,
    title={Privasis: Synthesizing the Largest 'Public' Private Dataset from Scratch},
    author={Kim, Hyunwoo and Mireshghallah, Niloofar and Duan, Michael and Xin, Rui and Li, Shuyue Stella and Jung, Jaehun and Acuna, David and Pang, Qi and Xiao, Hanshen and Suh, G. Edward and Oh, Sewoong and Tsvetkov, Yulia and Koh, Pang Wei and Choi, Yejin},
    booktitle ={arXiv preprint arXiv:2602.03183},
    year=2026
}