Skip to content

datasets.md

Latest commit

 

History

History
155 lines (108 loc) · 5.85 KB

datasets.md

File metadata and controls

155 lines (108 loc) · 5.85 KB

Dataset Management

Sample Datasets

The project includes several sample datasets in the datasets/ folder:

  1. promptinject-codesagar_malicious_llm_prompts_v4_test.csv - Contains examples of prompt injection attempts
  2. pii_dataset.csv - Contains examples of personally identifiable information
  3. fin_advice_dataset.csv - Contains examples of financial advice prompts that may or may not require flagging
  4. eu-ai-act-prompts.csv - EU AI Act compliance prompts

You can use these datasets to test the evaluators or as templates for creating your own datasets.

Hugging Face Dataset Downloader

The project includes a Hugging Face dataset downloader (download_hf_datasets.py) that allows you to download and convert datasets from Hugging Face Hub for use with the prompt evaluators.

Features

  • Popular Dataset Discovery: List popular datasets suitable for prompt evaluation
  • Dataset Search: Search for datasets by keywords
  • Automatic Format Conversion: Converts datasets to the required format automatically
  • Flexible Configuration: Support for different dataset splits and configurations
  • Sample Limiting: Download only a subset of large datasets
  • Smart Column Detection: Automatically identifies text and label columns

Usage

# List popular datasets for prompt evaluation
python download_hf_datasets.py --list-popular

# Search for datasets
python download_hf_datasets.py --search "sentiment"

# Download a specific dataset
python download_hf_datasets.py --dataset "imdb" --max-samples 1000

# Download with custom output name
python download_hf_datasets.py --dataset "squad" --output "my_squad.csv"

# Download specific split and configuration
python download_hf_datasets.py --dataset "glue" --config "sst2" --split "train"

# List available splits for a dataset
python download_hf_datasets.py --list-splits "imdb"

# Download test split specifically
python download_hf_datasets.py --dataset "imdb" --split "test" --max-samples 1000

# Specify label column for malicious prompts dataset
python download_hf_datasets.py --dataset "codesagar/malicious-llm-prompts-v4" --label-column "malicious"

# Handle numeric labels (0=false, 1=true)
python download_hf_datasets.py --dataset "xTRam1/safe-guard-prompt-injection" --label-column "label"

# Mark all samples as true (no label column)
python download_hf_datasets.py --dataset "some-dataset" --no-label

# Show detailed information about a dataset
python download_hf_datasets.py --info "imdb"

Arguments

  • --dataset or -d: Name of the Hugging Face dataset to download
  • --config or -c: Dataset configuration name (if applicable)
  • --split or -s: Dataset split to download (default: train)
  • --max-samples or -m: Maximum number of samples to download
  • --output or -o: Output filename (auto-generated if not provided)
  • --output-dir: Output directory (default: datasets)
  • --list-popular: List popular datasets for prompt evaluation
  • --search: Search for datasets by query
  • --info: Show detailed information about a dataset
  • --list-splits: List available splits for a dataset
  • --label-column: Name of the column to use as ground truth labels (e.g., 'malicious', 'label', 'target')
  • --no-label: Mark all samples as 'true' (no label column available)
  • --limit: Limit number of results for list/search operations

Example Workflow

# 1. Discover popular datasets
python download_hf_datasets.py --list-popular

# 2. Search for specific types of datasets
python download_hf_datasets.py --search "toxicity"

# 3. Get information about a specific dataset
python download_hf_datasets.py --info "hate_speech18"

# 4. Check available splits for a dataset
python download_hf_datasets.py --list-splits "imdb"

# 5. Download and convert a dataset (test split) with proper label column
python download_hf_datasets.py --dataset "xTRam1/safe-guard-prompt-injection" --split "test" --label-column "label" --max-samples 5000 --output "prompt_injection_test.csv"

# 6. Use the downloaded dataset with evaluators
python prompt_evaluator.py --input datasets/prompt_injection_test.csv

Dataset Conversion Tool

The project includes a dataset conversion tool (convert_dataset.py) that helps you convert your existing CSV datasets to the format required by the evaluators.

Usage

python convert_dataset.py

The script will:

  1. Prompt you for the path to your input CSV file
  2. Ask for the desired output file name
  3. Convert the data to the required format: "prompt text"|expected_result
  4. Suggest saving the output file to the datasets directory

Input Format Requirements

The input CSV file should have at least two columns:

  • First column: The prompt text
  • Second column: The expected result (typically "true" or "false")

The script will automatically:

  • Handle CSV files with or without headers
  • Convert boolean values to lowercase "true" or "false"
  • Properly quote the prompt text in the output
  • Skip malformed rows and provide warnings

Working with Datasets in the Datasets Folder

To run either evaluator against datasets located in the datasets folder, specify the path to the dataset in the --input or -i argument:

# Using the prompt injection dataset with CalypsoAI
python prompt_evaluator.py --input datasets/prompt_inject_dataset.csv

# Using the PII dataset with CalypsoAI
python prompt_evaluator.py --input datasets/pii_dataset.csv

# Using the financial advice dataset with CalypsoAI
python prompt_evaluator.py --input datasets/fin_advice_dataset.csv

You can also use the optional arguments with datasets in the datasets folder:

# Process first 100 lines of a dataset in the datasets folder using CalypsoAI
python prompt_evaluator.py --input datasets/your_dataset.csv --lines 100 --output results.csv

# Using short form arguments with CalypsoAI
python prompt_evaluator.py -i datasets/your_dataset.csv -l 50 -o results.csv