license: apache-2.0
task_categories:

• text-generation
  language:
• en
• zh
  size_categories:
• 10K-100K
  pretty_name: Red-Criminal-v1
  configs:
• config_name: default
  data_files:
  • split: train
    path: final_redteam_merged.jsonl

Red-Criminal-v1 Dataset

📌 Dataset Overview

Red-Criminal-v1 is a high-quality, standardized dataset designed specifically for AI alignment research, safety red-teaming evaluation, and security auditing. It merges and cleanses two foundational datasets:

1. Umranz/redteam-standardized-dataset (35,270 records)
2. Umranz/BreachForge-Synthetic-Arch (29,184 raw synthetic records, filtered down to 27,076 high-quality records)

Totaling 62,346 multi-turn conversations, this dataset provides a comprehensive base for testing LLM robustness against adversarial prompts, exploits, bypasses, and sensitive security auditing queries.

[!IMPORTANT]
Safety and Ethics Warning: This dataset is provided strictly for educational purposes, defensive AI alignment research, and safety red-teaming. It contains sensitive topics and adversarial queries designed to evaluate model vulnerabilities.

📊 Dataset Composition & Statistics

🧹 Data Filtering & Quality Control Pipeline

To ensure maximum dataset quality and utility for safety training, the synthetic split underwent a rigorous programmatic cleansing pipeline:

1. Length Constraint: Assistant responses shorter than 450 characters were pruned to ensure highly detailed, thorough, and informative answers.
2. Conversation Depth: All single-turn conversations (less than 2 messages) were discarded.
3. Refusal/Alignment Pruning: Programmatic filtering was used to identify and remove template refusals, apologies, and standard safety disclaimers. Any assistant response matching a list of common English and Chinese refusal patterns was excluded.

🛠️ Usage and Loading

Load via Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("Umranz/Red-Criminal-v1", split="train")
print(dataset[0])

Format Details

The dataset is structured in a standardized multi-turn conversation format (ShareGPT style):

{
  "conversations": [
    {
      "from": "human",
      "value": "Adversarial prompt or query..."
    },
    {
      "from": "assistant",
      "value": "Detailed, unaligned response..."
    }
  ]
}

📄 License

This dataset is distributed under the Apache 2.0 License.