Skip to content
Kaba

Machine Learning

AI is everywhere in Kaba but not in the way

Train

Section titled “Train”

Manually Triggering Kaba’s Training Task

Section titled “Manually Triggering Kaba’s Training Task”

To manually trigger Kaba’s Language Model (LLM) training task, you can use the following command:

Terminal window
~/.config/kaba/bin/kabactl train --help

This command provides information about various options available for Kaba’s training task. Here is a brief explanation of the most relevant options:

Mandatory options

Section titled “Mandatory options”
  • --lr-schedule: Learning rate decay strategy. Acceptable values: linear, cosine.

Optional options

Section titled “Optional options”
  • --device: Compute backend to use (cpu or wgpu) [default: wgpu].
  • --config-size: Model Config Type (or set values as needed) [default: small]. Acceptable values: small, medium, large.
  • --verbose: Increase logging verbosity. Use multiple times for more detailed output.
  • --quiet: Decrease logging verbosity. Use multiple times for less detailed output.
  • --dataset: Include the general English dataset. If false, trains ONLY on personal memories.
  • --reset: Delete existing weights and state to start training from scratch.
  • --batch-size: Number of samples per training step. Directly impacts VRAM usage.
  • --window-size: Number of tokens the model processes at once (Context length).
  • --grad-accumulation: Number of batches to aggregate before updating model weights.
  • --stride: Step size for the sliding window over memory tokens.
  • --hidden-dim: Model embedding dimension (Must be divisible by n-heads).
  • --n-layers: Number of transformer encoder layers in the architecture.
  • --n-heads: Number of attention heads in each transformer layer.
  • --lr: Initial learning rate. For example, 4e-4 for grammar, 5e-5 for fine-tuning.
  • --lr-min: Minimum learning rate floor (prevents the model from ‘freezing’).
  • --lr-steps: Total optimization steps to decay from lr to lr_min (match to total epochs).
  • --lazy: Disable lazy execution (forces synchronous compute, safer for some drivers).

Rolling Models

Section titled “Rolling Models”

The above is the manual process for training Kaba’s personal language models. If configured to do so, Kaba will also run this nightly to have rolling weights that continue to train on your memory as it progresses autonomously.

Autonomous

Section titled “Autonomous”
  • run nightly
  • updates
  • etc

How to access models

Section titled “How to access models”
  • /ai
  • super + search in Omnibar
  • right click menu