1 paper · 1 filter
Anurup Ganguli
Continually pre-training a large language model on heterogeneous text domains, without replay or task labels, has remained an unsolved architectural problem at LLM scale. Existing…