AR-to-MDLM conversions: 151M to 1.73B Collection 6 MDLM models converted from cooled AR parents in the AR suite, one per size, each trained for 50 tokens per non-embedding parameter. • 6 items • Updated 6 days ago
MDLM scaling suite: 59M to 1.73B Collection 31 cooled masked diffusion (MDLM) base models, same ladder and data as the AR suite, at several token budgets. Seeds on branches seed2, seed3. • 31 items • Updated 6 days ago
AR scaling suite: 59M to 1.73B Collection 33 cooled AR base models from one ladder and one data view, each size trained to several token budgets. Seeds on branches seed2, seed3. • 33 items • Updated 6 days ago