Design

Which solid notes this experiment tests, and how.

Results

20260901-095619-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag40_jean_zay-a8f7a80

commit: a8f7a80

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed1
condition_0_best_center_eval_accuracy0.765625
condition_0_center_eval_accuracy_change0.625
condition_0_center_train_accuracy_change0.8571428619325161
condition_0_condition_index0
condition_0_final_center_eval_accuracy0.75
condition_0_final_center_eval_parseable_rate1
condition_0_final_center_train_accuracy0.9047619104385376
condition_0_final_center_train_parseable_rate1
condition_0_initial_center_eval_accuracy0.125
condition_0_initial_center_eval_parseable_rate1
condition_0_initial_center_train_accuracy0.0476190485060215
condition_0_initial_center_train_parseable_rate1
condition_0_lr_direction0.5
condition_0_lr_magnitude4
condition_0_mean_fitness_std0.8956871708234151
condition_0_mean_perturbed_train_accuracy0.6978836437066396
condition_0_mean_perturbed_train_parseable_rate1
condition_0_mean_raw_score_std0.4396172026793162
condition_0_method_id1
condition_0_nonzero_reward_epoch_fraction1
condition_0_num_layers253
condition_0_rank1
condition_0_seed0
condition_0_shaped_epoch_fraction0
condition_0_sigma_direction0.001
condition_0_sigma_magnitude0.001
condition_0_tangent_project_direction0
decomposed_lr_direction0.5
decomposed_lr_magnitude4
decomposed_sigma_direction0.001
decomposed_sigma_magnitude0.001
decomposed_tangent_project_direction0
eggroll_lr1
eggroll_sigma0.001
epochs15
eval_batch_size8
eval_prompts128
generations_per_prompt6
max_new_tokens1024
method_decomposed_enabled1
method_plain_enabled0
noise_reuse4
num_center_eval_records4
num_completed_runs1
num_epoch_records15
num_methods1
num_ranks1
num_seeds1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
temperature1
top_k0
train_batch_size6
wall_clock_seconds35548.49023692799

Interpretation

Run 20260901-095619-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag40_jean_zay-a8f7a80
is a completed calibration point. It gives a usable end-to-end measurement for
the middle direction scale with the largest magnitude scale in this grid, and
therefore helps bound how aggressive the magnitude update can be before the
next repeated-seed test.

The boring alternative explanation remains fixed-batch fit rather than robust
reward learning. Treat this as one candidate scale, not evidence that the
decomposed parameterisation generalises beyond the fixed batch.