mini-AGI
An experimental language model trained from scratch on an 8GB VRAM GPU that keeps learning from what it reads.
SITE CAPTUREBased on information published on the service's website. Selection and review policy
WHAT IT DOES
What does it do?
mini-AGI stores weights in ordinary disk files and loads them onto the GPU when needed. During training, it expands its structure when capacity runs short and prunes unused parts. The provided description calls the current model a small experiment and says the final weights have not yet been released.
KEY FEATURES
Key features
- Train a model from scratch on a single 8GB VRAM GPU
- Take a training step for each chunk of the character stream
- Store weights on disk and load them onto the GPU when needed
- Expand capacity during training and prune unused parts