Design
Which solid notes this experiment tests, and how.
Results
20260902-082643-qwen3_4b_gsm8k_robustness_decomp025_05_bf16_seed1_jean_zay-9765cbb
commit: 9765cbb
| metric | value |
|---|---|
| center_eval_do_sample | 0 |
| center_eval_every | 5 |
| center_eval_temperature | 1 |
| center_eval_top_k | 0 |
| completed | 0 |
| decomposed_lr_direction | 0.25 |
| decomposed_lr_magnitude | 0.5 |
| 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 | 2 |
| num_completed_runs | 0 |
| num_epoch_records | 9 |
| 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 |
| torch_dtype_bfloat16 | 1 |
| torch_dtype_float16 | 0 |
| torch_dtype_float32 | 0 |
| train_batch_size | 6 |
| wall_clock_seconds | 10006.994043955114 |
20260902-082643-qwen3_4b_gsm8k_robustness_decomp025_05_bf16_seed1_jean_zay-9765cbb
commit: 9765cbb
| 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 | 0.7421875 |
| condition_0_center_eval_accuracy_change | 0.546875 |
| condition_0_center_train_accuracy_change | 0.5238095372915268 |
| condition_0_condition_index | 0 |
| condition_0_final_center_eval_accuracy | 0.7421875 |
| 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 | 0.1953125 |
| condition_0_initial_center_eval_parseable_rate | 1 |
| condition_0_initial_center_train_accuracy | 0.190476194024086 |
| condition_0_initial_center_train_parseable_rate | 1 |
| condition_0_lr_direction | 0.25 |
| condition_0_lr_magnitude | 0.5 |
| condition_0_mean_fitness_std | 0.9060774842898051 |
| condition_0_mean_perturbed_train_accuracy | 0.5888889253139495 |
| condition_0_mean_perturbed_train_parseable_rate | 1 |
| condition_0_mean_raw_score_std | 0.4710213641325633 |
| condition_0_method_id | 1 |
| condition_0_nonzero_reward_epoch_fraction | 1 |
| condition_0_num_layers | 253 |
| condition_0_rank | 1 |
| condition_0_seed | 1 |
| 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 | 0.25 |
| decomposed_lr_magnitude | 0.5 |
| 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 |
| torch_dtype_bfloat16 | 1 |
| torch_dtype_float16 | 0 |
| torch_dtype_float32 | 0 |
| train_batch_size | 6 |
| wall_clock_seconds | 16939.57706141402 |
Interpretation
Completed after the later fetch. This decomposed setting preserves the early reward-signal gain through the final center evaluation and is part of the strongest setting in the matched BF16 robustness batch.