Seed-twinned DataDecide models under one safety post-training recipe, with per-prompt judge labels. Paper: arXiv:2610.34074.
AI & ML interests
Open and Efficient LLM Pretraining
Recent Activity
View all activity
models 31
cerulean-works/castor-fineweb-pro-1B
Updated • 11
cerulean-works/castor-fineweb-edu-1B
Updated • 9
cerulean-works/castor-falcon-and-cc-qc-tulu-10p-1B
Updated • 9
cerulean-works/castor-falcon-and-cc-qc-orig-10p-1B
Updated • 7
cerulean-works/castor-falcon-and-cc-qc-20p-1B
Updated • 9
cerulean-works/castor-falcon-and-cc-qc-10p-1B
Updated • 9
cerulean-works/castor-falcon-and-cc-1B
Updated • 8
cerulean-works/castor-falcon-1B
Updated • 8 • 1
cerulean-works/castor-dolma1_7-no-reddit-1B
Updated • 11 • 1
cerulean-works/castor-dolma1_7-no-math-code-1B
Updated • 8