| {"seed_cosine":{"_note":"same method+lr, seed42 vs seed123; full vec-cosine, top-1 output dir u1, top-1 input dir v1, median top-8 principal angle (deg)","sft":{"5e-7":{"cos":0.066,"top1_u":0.988,"top1_v":0.074,"med_top8_angle_deg":26.6},"5e-6":{"cos":0.302,"top1_u":0.973,"top1_v":0.263,"med_top8_angle_deg":32.7},"5e-5":{"cos":0.36,"top1_u":0.901,"top1_v":0.403,"med_top8_angle_deg":44.4}},"rft":{"5e-7":{"cos":0.066,"top1_u":0.988,"top1_v":0.074,"med_top8_angle_deg":26.5},"5e-6":{"cos":0.3,"top1_u":0.975,"top1_v":0.26,"med_top8_angle_deg":32.6},"5e-5":{"cos":0.356,"top1_u":0.9,"top1_v":0.415,"med_top8_angle_deg":44.6}},"grpo":{"5e-7":{"cos":0.001,"top1_u":0.062,"top1_v":0.015,"med_top8_angle_deg":82.4,"_caveat":"dW norm 0.03 -> noise, not real orthogonality"}}},"frobenius_by_lr":{"sft":{"5e-7":1.016,"5e-6":2.86,"5e-5":9.21},"rft":{"5e-7":1.029,"5e-6":2.87,"5e-5":9.39},"grpo":{"5e-7":0.029}},"dapo_seed_cosine":{"5e-7":{"cos":0.005,"fro":0.033},"5e-6":{"cos":0.014,"fro":0.332},"5e-5":{"cos":0.01,"fro":3.515}},"dapo_lr_sensitivity":{"5e-7_to_5e-6":{"cos":0.381,"norm_ratio":10.03},"5e-6_to_5e-5":{"cos":0.125,"norm_ratio":10.58}},"lrs":["5e-7","5e-6","5e-5"]} |