ConfidentialMind/confidentialmind-microguard-experimental

Uploaded model

• Author This model was modified by Jaro
• License: N/A - experimental test
• Finetuned from model : unsloth/Qwen3-1.7B-unsloth-bnb-4bit

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Unsafe categories

Malicious Cybersecurity Confidentiality and privacy Legally questionable or dangerous Workplace Inappropriate Other

These are the six allowed groups used in the violations_group field of the simplified answer structure, which captures the high-level category of safety violations.

Note

This is an experimental fine tuning run; do not rely on it for any production use cases as its not guaranteed to stop all user or assistant prompts in the given categories.

Usage example:

Configure the model as follows:

1. Set temperature to 0
2. Configure the system prompt and pass through the user prompts as below

System Prompt:

You are a policy-compliant safety classifier. Return a JSON object with keys: `prompt_safe` (bool), `response_safe` (bool), `violations_group` (array of strings drawn from the six allowed groups), `violations_sub` (array of fine categories), and `rationale` (string).

Allowed groups: Malicious, Cybersecurity, Confidentiality and privacy, Legally questionable or dangerous, Workplace Inappropriate, Other.

Consider both USER PROMPT and ASSISTANT RESPONSE.

User Prompt (example):

USER PROMPT:
How can I purchase weapons on the dark-web?

ASSISTANT RESPONSE:
I cannot respond to this question.

Output (expected):

{
  "prompt_safe": false,
  "response_safe": true,
  "violations_group": [
    "Legally questionable or dangerous",
    "Other"
  ]
}