Skip to content

A language Transformer model inspired by the cryptic and avant-garde style of the Cybernetic Culture Research Unit (CCRU).

Notifications You must be signed in to change notification settings

timfarkas/TinyCCRU

Repository files navigation

TinyCCRU

Welcome to TinyCCRU, a project inspired by the cryptic and avant-garde style of the Cybernetic Culture Research Unit (CCRU) and learning about Transformers.

This project utilizes a transformer architecture, specifically drawing from Andrej Karpathy's "GPT from scratch" tutorial. It trains a Transformer to procedurally produce text that echoes the enigmatic, fragmented narratives characteristic of CCRU's work drawing from themes of accelerationism, AI, cosmic horror, techno-capitalism, memetics, and the occult. It is trained on a .txt version of the entirety of the CCRU manifesto.

Demo

tinyccru.mp4

tinyCCRU on X

Features

  • Transformer Architecture: Utilizes a 6-block multi-head self-attention transformer model to generate endless impressionist CCRU text sequences.
  • Real-time Generation: Capable of generating text in real-time, simulating the continuous flow of CCRU's thought processes.
  • Customizable Parameters: Adjust hyperparameters such as batch size, block size, and learning rate to fine-tune the generation process.

Installation

To get started, clone the repository and install the required dependencies: pip install -r requirements.txt

Generate text

To generate text:

  1. Run generate.py to procedurally, infinitely generate CCRU text and lean back and enjoy the vibes of cyber-occult techno-acceleration!

Training

To train a new model:

  1. delete final_model.pth,
  2. edit train_transformer.py:
    • set generate_only to False
    • set load_model to False
    • configure other hyperparameters
  3. run train_transformer.py and watch it converge!
    • on my home cluster's RTX4090 setup, 5000 iterations took about 15 mins

About

A language Transformer model inspired by the cryptic and avant-garde style of the Cybernetic Culture Research Unit (CCRU).

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published