reasoning-degeneration-dev/

prepretraining-eval-round4-v1

Round 4 (150M, 5B tokens): 4 conditions, seed=42, pre-SFT and post-SFT eval PPL. Front-load overtakes constant-mix on web PPL at this scale.

Rows: 8 Columns: 13

	Columns




	

Column Type Description

run_name Value('string') Unique run identifier: ppt_{scale}_{condition}_d{delta}_s{seed}

condition Value('string') Data scheduling condition: baseline, front-load, constant-mix, or anneal

scale Value('string')… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/prepretraining-eval-round4-v1.