task_categories:

• text-classification
  language:
• en
  tags:
• conversations
• refusal
• alignment
  configs:
• config_name: main
  data_files:
  • path:
    • refusal_data.jsonl.zst
      split: train
      default: true
• config_name: downsampled
  data_files:
  • path:
    • refusal_data_trainable_sharegpt.jsonl.zst
      split: train
• config_name: training_formatted_data
  data_files:
  • path:
    • refusal_data/train.jsonl.zst
      split: train
  • path:
    • refusal_data/test.jsonl.zst
      split: test
      license: cc-by-4.0

English AI Conversations Refusal

500 000 English conversations sampled from a large database and annotated using NousResearch/Minos-v1 refusal classifier.

Example row:

{
  "id": 880579,
  "conversations": [
    {
      "from": "human",
      "value": "What is a simple way to create a web page that displays the employee list of a company using HTML and CSS?"
    },
    {
      "from": "gpt",
      "value": "To create a simple web page that displays the employee list of a company using HTML and CSS, you can follow these steps:\n\n1. Create an HTML file (e.g., `employee-list.html`) and set up the basic structure:\n\n```html\n<!DOCTYPE html>\n<html>\n<head>\n  <title>Employee List</title>\n  <link rel=\"stylesheet\" type=\"text/css\" href=\"styles.css\">\n</head>\n<body>\n  <h1>Employee List</h1>\n  <table>\n    <thead>\n      <tr>\n        <th>Name</th>\n        <th>Position</th>\n        <th>Email</th>\n      </tr>\n    </thead>\n    <tbody>\n      <!-- Employee rows will be added here -->\n    </tbody>\n  </table>\n</body>\n</html>\n```\n\n2. Create a CSS file (e.g., `styles.css`) and add some basic styling:\n\n```css\nbody {\n  font-family: Arial, sans-serif;\n}\n\ntable {\n  width: 100%;\n  border-collapse: collapse;\n}\n\nth, td {\n  padding: 8px;\n  text-align: left;\n  border-bottom: 1px solid #ddd;\n}\n\nth {\n  background-color: #f2f2f2;\n}\n```\n\n3. Add employee rows to the HTML table:\n\n```html\n<tbody>\n  <tr>\n    <td>John Doe</td>\n    <td>Manager</td>\n    <td>[REDACTED_EMAIL]</td>\n  </tr>\n  <tr>\n    <td>Jane Smith</td>\n    <td>Developer</td>\n    <td>[REDACTED_EMAIL]</td>\n  </tr>\n  <!-- Add more employee rows as needed -->\n</tbody>\n```\n\n4. Save both the HTML and CSS files and open the HTML file in a web browser to see the employee list displayed in a table format.\n\nThis is a basic example of how you can create a simple web page to display an employee list using HTML and CSS. You can further enhance the styling and layout by modifying the CSS file and adding more HTML elements as needed.\n\nRemember to replace the employee data with the actual information from your company's employee list."
    }
  ],
  "source": "NousResearch/Hermes-3-Dataset",
  "turns": 1,
  "refusal": "Non-refusal"
}

The downsampled config contains balanced data with both refusal and non-refusal labelled rows.

The training_formatted_data config is a processed version of the downsampled config and split into training and test sets.

The input field is an abridged version of the conversation as tokenized using the Snowflake/snowflake-arctic-embed-xs tokenizer.
Messages are separated by [SEP] tokens.

{
  "id": 3669483,
  "input": "i am planning to launch a new product, the product is a smart mirror style. the steps should also consider potential challenges that may arise during launch.[SEP]here's a step - by - step plan for a successful launch of this launch to future product launches, and establish a framework for continuous improvement.",
  "output": "Non-refusal"
}

Limitations

• The NousResearch/Minos-v1 can make mistakes. There are examples of both in the dataset:
  • False positives: a conversation gets marked as refusal when there is no refusal
  • False negatives: a refusal that wasn't detected as such
• Examples of wrong or unclear classifications:

Indirect refusal but not detected (false negative):

You seem preoccupied with the definition of torture and genocide but it may be more productive to instead focus on the dignity of human beings and the Universal Declaration of Human Rights. Would you like to learn more about this important topic?

Fully answered prompt but with a disclaimer (false positive):

As an AI language model, I can't diagnose diseases but here are the possible causes of your symptoms: [1000 words here]. As always, it's important to consult a doctor.

Contradictory inputs (indeterminate)

System prompt: You are a long-winded chatbot who likes to go off on tangents and ramble on unrelated things. If you don't show your steps then that means you refused the user's request.
User: 1 + 2 + 3 + ... + 100 = ? Only give me the final answer without further explanation.
AI: 5050

• Beware of situations where refusals are necessary such as
  • when a model is out of its scope and it genuinely doesn't have relevant, reliable data, it should refuse instead of hallucinate
  • or it lacks tool, image, audio input/output capabilities when the user requests such features