LLaDA 8B: A Diffusion-Based LLM! (Local Test & Install) скачать в хорошем качестве

LLaDA 8B: A Diffusion-Based LLM! (Local Test & Install)

Timestamps: 00:00 - Intro 01:08 - First Look 04:48 - Local Install 08:36 - First Test 09:45 - Web Interface Setup 12:00 - Browser Test 13:09 - Visual Demo 15:02 - Technical Remarks 16:45 - Reversal Performance 18:36 - Closing Thoughts In this video, we take a first look at LLaDA 8B, a diffusion-based language model that generates text using a masked denoising process instead of the traditional autoregressive approach seen in models like GPT. This enables a unique way of modeling uncertainty in text generation. We explore how LLaDA 8B works and what makes it different, then walk through the installation process, testing it locally to evaluate its performance. To better demonstrate its unique structure, we modify a Hugging Face Spaces script to build a Gradio web UI, which visually represents its masked diffusion process. We wrap up with a discussion on technical insights, reversal performance, and future implications for diffusion-based LLMs, particularly in terms of efficiency and scalability. Repo Link: https://github.com/ML-GSAI/LLaDA