metadata
pipeline_tag: text-generation
RADD Small (lambda-dce)
This repository contains the small model checkpoint for RADD (Reparameterized Absorbing Discrete Diffusion), trained with the $\lambda$-DCE loss for 400k iterations.
RADD is a discrete diffusion model designed for language modeling that characterizes time-independent conditional probabilities. This approach allows for sampling acceleration via caching strategies and unifies absorbing discrete diffusion with any-order autoregressive models (AO-ARMs).
- Paper: Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data
- GitHub Repository: ML-GSAI/RADD
Usage
To use this model, you need to use the loading utility provided in the official repository:
from load_model import load_model
# Load the model and noise schedule
model, noise = load_model('JingyangOu/radd-lambda-dce', device='cuda')
For more details on sampling (e.g., using the DiffusionSampler or OrderedSampler), please refer to the scripts in the GitHub repository.
Citation
@misc{ou2024absorbingdiscretediffusionsecretly,
title={Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data},
author={Jingyang Ou and Shen Nie and Kaiwen Xue and Fengqi Zhu and Jiacheng Sun and Zhenguo Li and Chongxuan Li},
year={2024},
eprint={2406.03736},
archivePrefix={arXiv},
primaryClass={cs.LG},
}