Design

Which solid notes this experiment tests, and how.

Results

20260901-095408-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag10_jean_zay-b2aca31

commit: b2aca31

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_change0.640625
condition_0_center_train_accuracy_change0.7619047723710537
condition_0_condition_index0
condition_0_final_center_eval_accuracy0.765625
condition_0_final_center_eval_parseable_rate1
condition_0_final_center_train_accuracy0.8095238208770752
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_magnitude1
condition_0_mean_fitness_std0.9046175559361775
condition_0_mean_perturbed_train_accuracy0.6708995143572489
condition_0_mean_perturbed_train_parseable_rate1
condition_0_mean_raw_score_std0.4514849344889323
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_magnitude1
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_seconds35730.995360302

Interpretation

Run 20260901-095408-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag10_jean_zay-b2aca31
is a completed calibration point. It finished the planned epoch and center-eval
schedule under exact-answer reward only, so it belongs in the shortlist of
usable decomposed scales for a repeated-seed or fresh-batch follow-up.

The boring alternative explanation is fixed-batch overfitting. This run does
not rule that out, because the training prompts and deterministic eval slice
are still small. It should be compared with the other completed points first,
then repeated outside this fixed batch before supporting a method claim.