Language models
Design and train small language-model systems from scratch.
AI RESEARCH / VRA SERIES · ACTIVE RESEARCH
DVRA — DHAUL Vector Representation Architecture — researches and trains language models from first principles, beginning with the VRA series.
← Back to all projects
DVRA / DHAULAI RESEARCH / VRA SERIES
We start with small original models, run controlled experiments, study results, and improve through learning.
Design and train small language-model systems from scratch.
Study how text data is prepared, tokenized and presented to models.
Examine checkpoints, training results and model behavior.