Skip to content

AI Training #1

@childrda

Description

@childrda

💡 Note

This isn’t an issue with the system itself — it’s related to the AI model currently being used.

The model (llama3:8b) is a general-purpose language model and hasn’t been trained specifically for phishing or email classification. Because of that, it may:

Mislabel legitimate messages as phishing

Miss more subtle phishing attempts

🧠 Why This Happens

The model doesn’t yet understand the patterns unique to your organization’s emails (staff communications, newsletters, vendor notices, etc.).
To improve accuracy, it needs additional training data — real examples of both safe and phishing messages.

⚙️ Built-In Training Support

The system already supports data collection for fine-tuning.
When the following setting is enabled in config.json:

"train_ai": true

It will automatically:

Save each request sent to the AI classifier

Store the AI’s response in a training table

Those stored samples can later be reviewed and corrected. Once you have enough labeled examples, they can be used to fine-tune the model so it learns what a normal vs. phishing message looks like in your environment.

🚀 Next Steps

Enable "train_ai": true in your configuration.

Let the system collect several hundred or thousand labeled examples.

Use that dataset to fine-tune or retrain the model for higher accuracy.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions