{ "definition": "final residual before final RMSNorm, not logits, accuracy or LM quality", "changes": [ { "cohort": "legacy", "sample": "fresh-zh-1", "final_pre_norm_residual_relative_l2": 0.021898728857078685 }, { "cohort": "legacy", "sample": "fresh-zh-2", "final_pre_norm_residual_relative_l2": 0.03829561458389978 }, { "cohort": "legacy", "sample": "fresh-en-1", "final_pre_norm_residual_relative_l2": 0.02432907925877068 }, { "cohort": "legacy", "sample": "fresh-en-2", "final_pre_norm_residual_relative_l2": 0.034934933219615986 }, { "cohort": "gsm8k", "sample": "gsm8k-test-0000", "final_pre_norm_residual_relative_l2": 0.03901556767460501 }, { "cohort": "gsm8k", "sample": "gsm8k-test-0002", "final_pre_norm_residual_relative_l2": 0.04043223109300957 }, { "cohort": "gsm8k", "sample": "gsm8k-test-0003", "final_pre_norm_residual_relative_l2": 0.045740534396548564 }, { "cohort": "gsm8k", "sample": "gsm8k-test-0004", "final_pre_norm_residual_relative_l2": 0.04190792993325029 } ] }