tuandunghcmut/Primus-Seed-Conversation

Primus-Seed Conversational Dataset

A conversational dataset derived from trendmicro-ailab/Primus-Seed, designed for training language models as cybersecurity experts.

Dataset Description

This dataset converts the original Primus-Seed text completion dataset into a conversational format with 86,987 examples. Each example consists of a three-turn conversation between a system, user, and assistant focused on cybersecurity knowledge.

Dataset Statistics

• Total Examples: 86,987
• Format: Conversational (system/user/assistant messages)
• Domain: Cybersecurity, Cloud Security, Threat Detection, Vulnerability Management
• Language: English
• Size: ~328 MB (uncompressed), ~154 MB (download)

Dataset Structure

Data Fields

Each example contains:

• messages (list): A conversation with 3 messages:
  • role (string): One of "system", "user", or "assistant"
  • content (string): The message content
• source (string): Original data source identifier
• url (string): Source URL (when available)
• time (string): Timestamp of the original content

Data Example

{
  "messages": [
    {
      "role": "system",
      "content": "You are a cybersecurity expert with extensive knowledge in cloud security, threat detection, vulnerability management, and security best practices. Provide detailed, accurate, and actionable security guidance."
    },
    {
      "role": "user",
      "content": "What are the key security concepts and practices described here?"
    },
    {
      "role": "assistant",
      "content": "[Detailed security content from original dataset]"
    }
  ],
  "source": "cloudone-api-documentation",
  "url": "https://automation.trendmicro.com/cloudone/workload-security",
  "time": "2023-11-29 00:00:00"
}

Conversation Format

Each example follows this structure:

1. System Message: Establishes the AI as a cybersecurity expert with knowledge in:

   • Cloud security
   • Threat detection
   • Vulnerability management
   • Security best practices

2. User Message: Asks the model to explain security concepts using one of 10 varied question templates:

   • "Can you explain the security concepts and information covered in this document?"
   • "What are the key security concepts and practices described here?"
   • "Please provide a detailed explanation of the security information in this content."
   • And 7 other variations for diversity

3. Assistant Message: Contains the original security content from the Primus-Seed dataset

Use Cases

This dataset is ideal for:

• Fine-tuning language models as cybersecurity assistants
• Training conversational AI for security guidance
• Teaching models about cloud security and threat detection
• Creating security documentation Q&A systems
• Building AI security advisors

Dataset Creation

Source Data

The source dataset is trendmicro-ailab/Primus-Seed, which contains cybersecurity documentation and best practices from various sources including:

• Cloud security documentation
• API references
• Security guides
• Best practice documents

Conversion Process

1. Loaded the "default" subset, "train" split from Primus-Seed
2. For each example:
   • Created a system message establishing the cybersecurity expert persona
   • Generated a user question from 10 templates (randomly selected with seed=42)
   • Used the original "content" field as the assistant's response
3. Preserved all metadata (source, url, time)

Usage

Load with Datasets Library

from datasets import load_dataset

dataset = load_dataset("tuandunghcmut/Primus-Seed-Conversation")

Example Training with LLaMA-Factory

Add to your dataset_info.json:

{
  "primus_conversation": {
    "hf_hub_url": "tuandunghcmut/Primus-Seed-Conversation",
    "formatting": "sharegpt",
    "columns": {
      "messages": "messages"
    }
  }
}

Then use in training:

dataset: primus_conversation

Example with Transformers

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("tuandunghcmut/Primus-Seed-Conversation", split="train")
tokenizer = AutoTokenizer.from_pretrained("your-model")

# Access conversation
example = dataset[0]
for message in example["messages"]:
    print(f"{message['role']}: {message['content'][:100]}...")

Considerations for Use

Strengths

• High-quality cybersecurity content from trusted sources
• Conversational format ready for instruction tuning
• Diverse question templates for variety
• Comprehensive coverage of security topics
• Metadata preserved for filtering and analysis

Limitations

• All assistant responses are factual documentation (not creative generation)
• User questions follow template patterns (10 variations)
• Single-turn conversations only
• English language only
• Focused on cybersecurity domain

License

Please refer to the original Primus-Seed dataset for licensing information.

Citation

If you use this dataset, please cite the original Primus-Seed dataset:

@dataset{primus_seed,
  title={Primus-Seed},
  author={Trend Micro AI Lab},
  year={2023},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/trendmicro-ailab/Primus-Seed}
}

Dataset Card Authors

• tuandunghcmut

Acknowledgments

Special thanks to Trend Micro AI Lab for creating and sharing the original Primus-Seed dataset.