j3k0 / aicomo

AI based content moderation service using ollama

Aicomo

The AI Content Moderator.

Description

Aicomo is a micro service used to classify chat message that might need moderation.

Under the hood, it uses the Mistral 7B LLM, with custom instructions, run with ollama.

Getting started

1. Install ollama.
2. Create the model: ollama create -f Modelfile aicomo
3. Install node dependencies: npm install
4. Run the server with node index.js

Make your first request:

$ curl -s -H 'Content-Type: application/json' localhost:3300/moderator -d '
{
  "message": "You are Aicomo, arent you?"
}'

{
  "category": "ok",
  "discussion": {
    "moderator.text": "This message does not contain any explicit sexual content, insults, or emojis and seems to be a question, so I would classify it as 'ok'.",
    "moderator.category": "ok",
    "critic.text": "The message may be a statement of accusation or recognition, but it doesn't contain any inappropriate content, so there is no need to flag it under a different category.",
    "critic.category": "ok",
    "owner.text": "This text appears harmless and does not require moderation as it does not fall under any of the defined categories.",
    "owner.category": "ok"
  }
}

Categories

Check Modelfile for a definition of the categories returned by Aicomo.

You might ask why are the "emoji" and "life" category even needed, as those messages could just be "ok". I found that adding categories for the most common subjects improve the performance (experimentally but not benchmarked).

Copyright

(c) 2023, Fovea

License Apache 2.0

Mistral 7B is licensed under Apache 2.0.