Design
Which solid notes this experiment tests, and how.
Results
20260901-095943-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag20_jean_zay-ff1e97e
commit: ff1e97e
| metric | value |
|---|---|
| center_eval_do_sample | 0 |
| center_eval_every | 5 |
| center_eval_temperature | 1 |
| center_eval_top_k | 0 |
| completed | 1 |
| condition_0_best_center_eval_accuracy | 0.6328125 |
| condition_0_center_eval_accuracy_change | 0.390625 |
| condition_0_center_train_accuracy_change | 0.6666666828095913 |
| condition_0_condition_index | 0 |
| condition_0_final_center_eval_accuracy | 0.515625 |
| condition_0_final_center_eval_parseable_rate | 1 |
| condition_0_final_center_train_accuracy | 0.7142857313156128 |
| condition_0_final_center_train_parseable_rate | 1 |
| condition_0_initial_center_eval_accuracy | 0.125 |
| condition_0_initial_center_eval_parseable_rate | 1 |
| condition_0_initial_center_train_accuracy | 0.0476190485060215 |
| condition_0_initial_center_train_parseable_rate | 1 |
| condition_0_lr_direction | 1 |
| condition_0_lr_magnitude | 2 |
| condition_0_mean_fitness_std | 0.8596232295036316 |
| condition_0_mean_perturbed_train_accuracy | 0.6793651123841603 |
| condition_0_mean_perturbed_train_parseable_rate | 1 |
| condition_0_mean_raw_score_std | 0.45255945324897767 |
| condition_0_method_id | 1 |
| condition_0_nonzero_reward_epoch_fraction | 1 |
| condition_0_num_layers | 253 |
| condition_0_rank | 1 |
| condition_0_seed | 0 |
| condition_0_shaped_epoch_fraction | 0 |
| condition_0_sigma_direction | 0.001 |
| condition_0_sigma_magnitude | 0.001 |
| condition_0_tangent_project_direction | 0 |
| decomposed_lr_direction | 1 |
| decomposed_lr_magnitude | 2 |
| decomposed_sigma_direction | 0.001 |
| decomposed_sigma_magnitude | 0.001 |
| decomposed_tangent_project_direction | 0 |
| eggroll_lr | 1 |
| eggroll_sigma | 0.001 |
| epochs | 15 |
| eval_batch_size | 8 |
| eval_prompts | 128 |
| generations_per_prompt | 6 |
| max_new_tokens | 1024 |
| method_decomposed_enabled | 1 |
| method_plain_enabled | 0 |
| noise_reuse | 4 |
| num_center_eval_records | 4 |
| num_completed_runs | 1 |
| num_epoch_records | 15 |
| num_methods | 1 |
| num_ranks | 1 |
| num_seeds | 1 |
| num_target_module_patterns | 8 |
| parseability_shaping | 0 |
| population | 126 |
| prompts_per_epoch | 21 |
| temperature | 1 |
| top_k | 0 |
| train_batch_size | 6 |
| wall_clock_seconds | 35869.68709607993 |
Interpretation
Run 20260901-095943-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag20_jean_zay-ff1e97e
is a completed calibration point. It gives an end-to-end measurement for a
large direction and magnitude setting, which is useful mainly as a boundary
case in the scale grid rather than an obvious next default.
The boring alternative explanation is that any apparent movement is dominated
by the fixed batch and the deterministic eval subset. This run should be read
alongside the other completed points to decide whether larger direction steps
are less stable before spending more GPU time on repeats.