Long reads, experiments and insights. Written by the people who build Rapidata — for the people who train models.
Existing approaches for aligning to human preferences typically make one of two compromises. Methods that directly optimize on human preferences, such as Diffusion-DPO [5],…
Mads Kuhlmann-JoergensenSep 30, 20268 min
Collecting human data at scale usually means going through a platform that connects you to a paid crowd. Prolific has become the industry standard for this kind of work, with a…
Elias TeikariAug 3, 20265 min
In the past few years, text-to-image models have evolved from DALL-E [7] to Stable Diffusion [8] to more recently Imagen 4 [9]. Early diffusion-based models struggled with simple…
Daniil PyatkoJul 9, 20259 minWhat we’re learning from human feedback at scale.
Reinforcement Learning from Human Feedback is most commonly associated with the final training stages of large language models like ChatGPT or Claude. It’s what helps these models…
Daniil PyatkoMay 12, 20255 minTL;DR: We collected 1.5 million annotations from >150 thousand individual humans using Rapidata via the Python API to build a dataset of detailed human feedback for text-to-image…
Mads Kuhlmann-JoergensenJan 10, 20256 minText-to-image models like DALL-E 3, Stable Diffusion, MidJourney, and Flux.1 have gained popularity for generating images from text prompts. However, benchmarking these models is…
Mads Kuhlmann-JoergensenSep 12, 20249 minObject detection in computer vision is a fascinating blend of mathematics, algorithms, and machine learning that allows computers to identify and locate objects within an image or…
Marian KannwischerApr 30, 20243 min