Design

Which solid notes this experiment tests, and how.

Results

20260903-091656-qwen3_4b_gsm8k_first_final_plain_jean_zay-0fe35ae

commit: 0fe35ae

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_change0.59375
condition_0_center_train_accuracy_change0.8095238134264946
condition_0_condition_index0
condition_0_final_center_eval_accuracy0.7109375
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_accuracy0.1171875
condition_0_initial_center_eval_parseable_rate1
condition_0_initial_center_train_accuracy0.095238097012043
condition_0_initial_center_train_parseable_rate1
condition_0_lr1
condition_0_mean_fitness_std0.8499476075172424
condition_0_mean_perturbed_train_accuracy0.7306878725687663
condition_0_mean_perturbed_train_parseable_rate1
condition_0_mean_raw_score_std0.4276362915833791
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_direction0.25
decomposed_lr_magnitude0.5
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
grpo_lora_alpha1
grpo_lora_init_std0.01
grpo_lr1e-05
grpo_weight_decay0
max_new_tokens1024
method_decomposed_enabled0
method_grpo_enabled0
method_plain_enabled1
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
torch_dtype_bfloat161
torch_dtype_float160
torch_dtype_float320
train_batch_size6
wall_clock_seconds16476.463732297998

Interpretation