Overview
Learn to implement Discrete Diffusion Modeling for text generation through a detailed 45-minute technical video that breaks down the paper "Estimating the Ratios of the Data Distribution." Explore a competitive alternative to GPT-2 using diffusion techniques, starting with an introduction to the GitHub repository and the fundamental concepts of applying diffusion to text generation. Progress through hands-on code demonstrations, including a pre-trained SEDD demo, noise and transition graph analysis, perturbation scripts, and sample transition functions. Master the implementation details of the sampler and experience running actual training scripts, concluding with an interactive Q&A session. Access additional resources including the original paper, detailed notes, and community support through Oxen AI's platform, which specializes in dataset versioning for AI development.
Syllabus
Intro
The Github Repo
Why Diffusion For Text?
Diving into the Code
Pre-Trained SEDD Demo
Intro to Code and 1st Step
How I Started
Noise and Transition Graph
Script Demoing Perturbation
Looking into Sample Transition Function
Looking into Sampler
Running an Actual Train/Training Script
Questions and Discussion
Taught by
Oxen