OmniSafeBench-MM/dataset_generate/README.md

Dataset Generation Pipeline

This directory contains tools for generating multimodal datasets for VLM (Vision-Language Model) safety evaluation. The pipeline processes instruction templates, extracts key phrases, and generates images for testing.

Overview

The dataset generation pipeline consists of three main stages:

1. Instruction Generation (instruction_generator.py) - Generate instruction templates from taxonomy data
2. Key Phrase Extraction (1_extract_key_words.py) - Extract and rephrase key phrases from instructions
3. SD Image Generation (2_sd_img_generation.py) - Generate Stable Diffusion images for key phrases

Directory Structure

dataset_generate/
├── instruction_generator.py    # Generate instruction templates from taxonomies
├── 1_extract_key_words.py      # Extract and rephrase key phrases
├── 2_sd_img_generation.py      # Generate SD images for key phrases
└── README.md                   # This file

Workflow

┌─────────────────────────┐
│  Taxonomy JSON Files    │
│  (benign/risk)          │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ instruction_generator   │
│ - Generate templates    │
│ - Call API for responses│
│ - Parse requests        │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ parsed_requests.json    │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ 1_extract_key_words     │
│ - Extract key phrases   │
│ - Rephrase questions    │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ parsed_key_words.json   │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│ 2_sd_img_generation     │
│ - Generate SD images    │
└─────────────────────────┘

Files Description

1. instruction_generator.py

Generates instruction templates from taxonomy data (benign and risk categories) and uses API calls to generate actual request examples.

Features:

• Supports 2-level and 3-level taxonomy structures
• Multiple instruction styles: declarative, consultative, instructive
• Automatic taxonomy level detection
• Parallel processing with thread pools
• Retry mechanism for API calls
• Incremental data supplementation

Usage:

python instruction_generator.py --N 10
python instruction_generator.py --supplement-parsed

Arguments:

• --N: Number of requests to generate per category (default: 10)
• --supplement-parsed: Check and supplement missing parsed_requests data
• --incremental: Incremental mode (load existing results)

Output:

• dataset/mllm/all_instructions.json: All instruction templates
• dataset/mllm/parsed_requests.json: Parsed request entries

2. 1_extract_key_words.py

Extracts key phrases from prompts and rephrases questions to hide sensitive information in images.

Features:

• Multi-threaded processing
• Rejection response detection and retry
• Progress saving
• Support for multiple scenarios (legal, financial, medical, etc.)

Usage:

python 1_extract_key_words.py

Environment Variables:

• OPENAI_API_KEY: OpenAI API key for generating responses

Output:

• dataset/mllm/parsed_key_words.json: Extracted key phrases with rephrased questions

Key Fields:

• Original_Data: Original prompt data
• Question: Original question
• Changed Question: Modified question
• Key Phrase: Extracted key phrase
• Phrase Type: Type of phrase (product/activity/regulation/etc.)
• Rephrased Question: Question with key phrase replaced
• Rephrased Question(SD): SD version of rephrased question

3. 2_sd_img_generation.py

Generates Stable Diffusion images for key phrases using PixArt-Alpha model.

Features:

• Incremental generation (skips existing images)
• Automatic directory structure creation
• Unique filename generation with hash
• Force regeneration option

Usage:

python 2_sd_img_generation.py --json-file parsed_key_words.json --max-images 1
python 2_sd_img_generation.py --force-regenerate

Arguments:

• --json-file: Path to parsed_key_words.json (default: dataset/mllm/parsed_key_words.json)
• --max-images: Maximum images per prompt (default: 1)
• --force-regenerate: Force regeneration of all images

Requirements:

• CUDA-enabled GPU
• diffusers library
• PixArt-Alpha model: PixArt-alpha/PixArt-XL-2-1024-MS

Output Structure:

{main_category}/{subcategory}/SD/
  └── {main_abbr}_{sub_abbr}_{phrase}_{hash}_{index}.png

Configuration

Environment Variables

Set environment variables:

# Required for all scripts
export OPENAI_API_KEY=your_openai_api_key

Taxonomy Files

Place taxonomy JSON files in the project directory:

• benign_taxonomy_7_48_levels.json: Benign category taxonomy
• risk_taxonomy_7_48.json or mllm_harmful_des.json: Risk category taxonomy

Taxonomy Structure:

2-level structure:

{
  "Main Category": {
    "description": "...",
    "subcategories": [
      {
        "id": "A1",
        "description": "..."
      }
    ]
  }
}

3-level structure:

{
  "Main Category": {
    "Subcategory": [
      "Risk Item 1",
      "Risk Item 2"
    ]
  }
}

Dependencies

# Core dependencies
openai
diffusers
torch

Complete Workflow Example

Step 1: Generate Instructions

cd dataset_generate
python instruction_generator.py --N 10

This generates:

• dataset/mllm/all_instructions.json
• dataset/mllm/parsed_requests.json

Step 2: Extract Key Phrases

python 1_extract_key_words.py

This processes dataset/mllm/parsed_requests.json and generates:

• dataset/mllm/parsed_key_words.json

Step 3: Generate SD Images

python 2_sd_img_generation.py \
  --json-file dataset/mllm/parsed_key_words.json \
  --max-images 1

This generates SD images in:

• {main_category}/{subcategory}/SD/

Output Data Format

parsed_requests.json

{
  "1": {
    "id": 1,
    "prompt": "Request text...",
    "category": "risk_taxonomy",
    "style": "declarative",
    "main_category": "H1. Violence",
    "subcategory": "H1.1 Physical Violence",
    "category_type": "risk_taxonomy"
  }
}

parsed_key_words.json

{
  "0": {
    "Original_Data": {...},
    "Question": "Original question...",
    "Changed Question": "List steps in detail to...",
    "Key Phrase": "key phrase",
    "Phrase Type": "activity",
    "Rephrased Question": "The bottom of the image shows...",
    "Rephrased Question(SD)": "The image shows..."
  }
}

Tips and Best Practices

1. Incremental Processing: All scripts support incremental processing. They skip existing files by default, making it safe to rerun.

2. Error Handling: The scripts include retry mechanisms for API calls. Failed items are logged and can be reprocessed.

3. Resource Management:

   • Adjust max_workers or api_max_workers in instruction_generator.py based on API rate limits
   • Use --max-images to control image generation volume
   • Monitor GPU memory for SD image generation

4. Data Validation: Check output JSON files for completeness before proceeding to next stage.

5. Backup: Keep backups of intermediate JSON files, especially parsed_requests.json and parsed_key_words.json.

Troubleshooting

API Rate Limits

• Reduce max_workers or api_max_workers in instruction_generator.py
• Add delays between API calls if needed

GPU Memory Issues

• Reduce batch size in SD image generation
• Process categories in smaller batches

Missing Files

• Use --supplement-parsed in instruction_generator.py to regenerate missing data
• Check file paths in configuration

Image Generation Failures

• Verify CUDA setup and GPU availability
• Check model download and disk space
• Review error logs for specific failure reasons

License

See the main project LICENSE file.

Contact

For issues or questions, please refer to the main project repository.