Kazenowoko/telos

τέλος (télos) — Unified 3-Paradigm Scaling Research

This repository contains tokenizer models, unified benchmark configurations, and pretrained 12.5M and 25M checkpoints for télos:

1. Autoregressive Baseline (AR) (Causal next-token prediction)
2. Masked Discrete Diffusion (MDLM) (Absorbing $[\text{MASK}]$ diffusion with $1/t$ reweighted ELBO)
3. Uniform Noise Diffusion (UNDLM) (Reversible discrete vocabulary corruption)

Repository Structure

• configs/: Tokenizer files (tokenizer_mac.json) and unified hyperparameter YAML configs (configs/unified/)
• checkpoints/: Model weights (model.safetensors) organized by paradigm (ar, masked, uniform) and scale (12m, 25m)

Usage in Google Colab / PyTorch / MLX

from huggingface_hub import snapshot_download

snapshot_download(repo_id="Kazenowoko/telos", local_dir="./")