The blog post discusses the author's personal project of training small language models (LLMs) to play the game 'Connections' more effectively than larger frontier models. The author outlines the criteria for selecting a suitable ...
About:
Brian Fitzgerald is the author of the website brianfitzgerald.xyz, which is a blog focused on topics related to training reasoning models, post-training techniques for open LLMs, curriculum learning, and synthetic data for LLMs. The blog also covers advancements in diffusion models and video diffusion techniques.
Website:
Specializations:
Outgoing Links:
Subscribe to RSS:
The text discusses the training of reasoning models using Direct Preference Optimization (DPO) as a cheaper alternative to the current state of the art. It reviews various papers on the topic and discusses the limitations of DPO a...
The text is a review of post-training techniques from recent open LLMs. It discusses pivotal token search, GRPO, and Online Merging Optimizer. The author provides a brief overview of each technique and its application in large-sca...
The author discusses the use of DPO and Logit Pruning for fine-tuning LLMs for code generation tasks. They experimented with different techniques and datasets, and evaluated the performance using MBPP and HumanEval. The results sh...
The author has created a synthetic dataset for training LLMs to use tools and is training a model using DPO to improve accuracy. The author is training a model for tool use, entirely with synthetic data, using DPO to improve accur...
The text discusses the development of a 77M T5 model to expand prompts, aiming to improve the quality and prompt alignment of the generated images. Various approaches were tried, including CLIP augmentation, U-Net augmentation, Ti...
0Understanding StyleAligned | Brian Fitzgerald
2024-01-28 •
The text discusses the limitations of current generative models and introduces two new techniques, StyleAligned and ConsisStory, that aim to accurately capture the style of a given image or the identity of a given person with mini...
The text discusses the development of Stable Video Diffusion (SVD) and its ability to generate longer videos. It explains the use of positional embeddings in SVD and the application of the YaRN technique to generate longer sequenc...