Design

This diagnostic tests the draft scaling suspicion raised after the grouped
GSM8K reward follow-ups: raw decomposed direction and magnitude learning rates
may not correspond to comparable effective weight-space updates. It is not an
accuracy sweep. The run uses one grouped GSM8K reward batch, applies one ES
update, and measures the resulting parameter-space and effective DoRA
weight-space update sizes for plain EGGROLL and decomposed EGGROLL.

Results

20260803-104230-pretrained_llm_gsm8k_scale_diagnostic_jean_zay-cb7f7a9

commit: cb7f7a9

metricvalue
decomposed_lr_direction0.02
decomposed_lr_magnitude_0_mean_effective_weight_relative_update0.0008017193681629564
decomposed_lr_magnitude_0_mean_magnitude_effective_weight_delta_norm0
decomposed_lr_magnitude_0_mean_magnitude_param_relative_update0
decomposed_lr_magnitude_0_mean_magnitude_to_direction_effective_weight_delta_rationull
decomposed_lr_magnitude_0_value0
decomposed_lr_magnitude_1_mean_effective_weight_relative_update0.0008040715523299729
decomposed_lr_magnitude_1_mean_magnitude_effective_weight_delta_norm0.12167434429702485
decomposed_lr_magnitude_1_mean_magnitude_param_relative_update6.145854126400007e-05
decomposed_lr_magnitude_1_mean_magnitude_to_direction_effective_weight_delta_rationull
decomposed_lr_magnitude_1_value0.02
decomposed_lr_magnitude_2_mean_effective_weight_relative_update0.0008585934467492261
decomposed_lr_magnitude_2_mean_magnitude_effective_weight_delta_norm0.608371534614009
decomposed_lr_magnitude_2_mean_magnitude_param_relative_update0.00030729263619639946
decomposed_lr_magnitude_2_mean_magnitude_to_direction_effective_weight_delta_rationull
decomposed_lr_magnitude_2_value0.1
decomposed_mean_direction_effective_weight_delta_norm1.5872272590950383
decomposed_mean_direction_param_relative_update0.0008023099603004136
decomposed_mean_effective_weight_relative_update0.0008214614557473851
decomposed_mean_fitness_std0.7432134648164114
decomposed_mean_magnitude_effective_weight_delta_norm0.24334862630367793
decomposed_mean_magnitude_param_relative_update0.00012291705915346652
decomposed_mean_magnitude_to_direction_effective_weight_delta_rationull
decomposed_mean_magnitude_to_direction_param_update_rationull
decomposed_mean_train_accuracy0.020833333333333332
decomposed_mean_train_parseable_rate0.9895833333333334
decomposed_mean_used_parseability_shaping0.5
decomposed_sigma_direction0.01
decomposed_sigma_magnitude0.02
decomposed_tangent_project_direction0
eggroll_lr0.02
eggroll_mean_fitness_std0.6009588241577148
eggroll_mean_train_accuracy0.0078125
eggroll_mean_train_parseable_rate0.9869791666666666
eggroll_mean_used_parseability_shaping0.6666666666666666
eggroll_mean_weight_relative_update0.0006769713713991269
eggroll_sigma0.01
generations_per_prompt8
num_decomposed_lr_magnitudes3
num_ranks2
num_seeds3
num_target_modules7
population64
prompts_per_epoch8
wall_clock_seconds2001.3164142081514

Interpretation

Run 20260803-104230-pretrained_llm_gsm8k_scale_diagnostic_jean_zay-cb7f7a9
supports the scale-mismatch explanation. With the old equal decomposed learning
rates, lr_magnitude=0.02 produced a mean magnitude effective-weight delta of
0.12167434429702485, while the mean direction effective-weight delta was
1.5872272590950383; the ratio of mean deltas is therefore about
0.07665842657364502. Raising lr_magnitude to 0.1 raised the magnitude
delta to 0.608371534614009, or about 0.38329201513390937 of the direction
delta. That makes the previous observation that 0.1 was the least bad
magnitude LR much less weird: 0.02 was probably under-scaling magnitude
relative to direction.

The boring alternative explanation is still that the GSM8K reward batch is
mostly sparse/noisy. The diagnostic does not rule that out: train accuracy was
low, parseability shaping was used on a substantial fraction of batches, and
this was only a one-step measurement. What it does rule against is the idea
that equal raw direction/magnitude learning rates were already a fair scale
comparison.

Next, do not run another broad accuracy sweep with equal decomposed rates.
Either tune magnitude to match a target fraction of the direction effective
delta, or run a small two-dimensional scale grid over direction and magnitude
where the reported metric is both reward movement and effective update size.