Design

Which solid notes this experiment tests, and how.

Results

20260831-203342-qwen3_4b_gsm8k_fixed_batch_method_comparison_plain_only_jean_zay-5bd4dfe

commit: 5bd4dfe

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_change0.3671875
condition_0_center_train_accuracy_change0.6666666828095913
condition_0_condition_index0
condition_0_final_center_eval_accuracy0.4921875
condition_0_final_center_eval_parseable_rate1
condition_0_final_center_train_accuracy0.7142857313156128
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_lr1
condition_0_mean_fitness_std0.8663099497556687
condition_0_mean_perturbed_train_accuracy0.7007936924695969
condition_0_mean_perturbed_train_parseable_rate1
condition_0_mean_raw_score_std0.4381131261587143
condition_0_method_id0
condition_0_nonzero_reward_epoch_fraction1
condition_0_num_layers253
condition_0_rank1
condition_0_seed0
condition_0_shaped_epoch_fraction0
decomposed_lr_direction1
decomposed_lr_magnitude1
decomposed_sigma_direction0.001
decomposed_sigma_magnitude0.001
decomposed_tangent_project_direction0
eggroll_lr1
eggroll_sigma0.001
epochs20
eval_batch_size8
eval_prompts128
generations_per_prompt6
max_new_tokens1024
method_decomposed_enabled0
method_plain_enabled1
noise_reuse4
num_center_eval_records5
num_completed_runs1
num_epoch_records20
num_methods1
num_ranks1
num_seeds1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
temperature1
top_k0
train_batch_size6
wall_clock_seconds24683.30508492398

Interpretation