same-data-different-losses / _src /seed_lr_sensitivity.json
AlexWortega's picture
Interactive blog: Same Data, Different Losses, Same Circuits?
378a074 verified
Raw
History Blame Contribute Delete
3.5 kB
{
"_about": "Seed + learning-rate sensitivity extension (Qwen3-4B-Instruct-2507, attention-only LoRA r32 a64). Two seeds {42,123} x three peak LRs {5e-7,5e-6,5e-5} per method. Cosine is of the gauge-invariant stacked LoRA delta dW=(alpha/r)BA over 144 modules. 'u1'/'v1' = top-1 left/right singular direction. GRPO rows at lr5e-6/5e-5 pending full training.",
"frobenius_norm": {
"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
}
},
"within_config_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"
}
}
},
"cross_method_cosine_matched_lr_seed": {
"_note": "same lr+seed, mean over seeds; contrast for the within-seed comparison",
"sft__rft": {
"5e-7": 0.996,
"5e-6": 0.978,
"5e-5": 0.836
},
"sft__grpo": {
"5e-7": 0.022
},
"rft__grpo": {
"5e-7": 0.021
}
},
"cross_method_same_seed42_subspace": {
"sft__rft": {
"5e-7": {
"cos": 0.996,
"top1_u": 0.999,
"top1_v": 1.0,
"med_top8_angle_deg": 3.7
},
"5e-6": {
"cos": 0.976,
"top1_u": 0.985,
"top1_v": 0.999,
"med_top8_angle_deg": 11.9
},
"5e-5": {
"cos": 0.831,
"top1_u": 0.903,
"top1_v": 0.947,
"med_top8_angle_deg": 26.7
}
}
},
"lr_sensitivity_fixed_seed42": {
"_note": "fixed method+seed, vary lr; norm_ratio vs lr_ratio tests 'rescale vs rotate'. norm/lr ~1 would mean pure rescale.",
"sft": {
"5e-7_to_5e-6": {
"cos": 0.544,
"norm_ratio": 2.79,
"lr_ratio": 10,
"norm_over_lr": 0.28
},
"5e-6_to_5e-5": {
"cos": 0.557,
"norm_ratio": 3.23,
"lr_ratio": 10,
"norm_over_lr": 0.32
}
}
},
"random_baseline": {
"sft5e-5_s42__rft5e-7_s123": {
"cos": 0.027,
"top1_u": 0.273,
"med_top8_angle_deg": 75.6
}
},
"seed_lmc_sft_lr5e-6": {
"_note": "masked-answer CE on GSM8K, alpha*s42+(1-a)*s123, n=24",
"curve": {
"0.0": 3.1878,
"0.25": 3.1773,
"0.5": 3.1646,
"0.75": 3.1499,
"1.0": 3.1333
},
"midpoint_barrier": 0.0041,
"verdict": "no barrier -> same basin"
}
}