Design
Which solid notes this experiment tests, and how.
Results
20260803-161316-pretrained_llm_gsm8k_scale_confirm_jean_zay-ef04761
commit: ef04761
| metric | value |
|---|---|
| both_success_rate | 0.3333333333333333 |
| decomposed_grid_0_lr_direction | 0.005 |
| decomposed_grid_0_lr_magnitude | 0.2 |
| decomposed_grid_0_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_0_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_0_success_rate | 0.3333333333333333 |
| decomposed_grid_0_win_rate | 0.3333333333333333 |
| decomposed_grid_10_lr_direction | 0.015 |
| decomposed_grid_10_lr_magnitude | 0.3 |
| decomposed_grid_10_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_grid_10_mean_final_accuracy | 0.027777778605620067 |
| decomposed_grid_10_success_rate | 0.3333333333333333 |
| decomposed_grid_10_win_rate | 0.3333333333333333 |
| decomposed_grid_11_lr_direction | 0.015 |
| decomposed_grid_11_lr_magnitude | 0.35 |
| decomposed_grid_11_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_grid_11_mean_final_accuracy | 0.027777778605620067 |
| decomposed_grid_11_success_rate | 0.3333333333333333 |
| decomposed_grid_11_win_rate | 0.3333333333333333 |
| decomposed_grid_1_lr_direction | 0.005 |
| decomposed_grid_1_lr_magnitude | 0.25 |
| decomposed_grid_1_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_1_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_1_success_rate | 0.3333333333333333 |
| decomposed_grid_1_win_rate | 0.3333333333333333 |
| decomposed_grid_2_lr_direction | 0.005 |
| decomposed_grid_2_lr_magnitude | 0.3 |
| decomposed_grid_2_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_2_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_2_success_rate | 0.3333333333333333 |
| decomposed_grid_2_win_rate | 0.3333333333333333 |
| decomposed_grid_3_lr_direction | 0.005 |
| decomposed_grid_3_lr_magnitude | 0.35 |
| decomposed_grid_3_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_3_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_3_success_rate | 0.3333333333333333 |
| decomposed_grid_3_win_rate | 0.3333333333333333 |
| decomposed_grid_4_lr_direction | 0.01 |
| decomposed_grid_4_lr_magnitude | 0.2 |
| decomposed_grid_4_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_grid_4_mean_final_accuracy | 0.027777778605620067 |
| decomposed_grid_4_success_rate | 0.3333333333333333 |
| decomposed_grid_4_win_rate | 0.3333333333333333 |
| decomposed_grid_5_lr_direction | 0.01 |
| decomposed_grid_5_lr_magnitude | 0.25 |
| decomposed_grid_5_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_5_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_5_success_rate | 0.3333333333333333 |
| decomposed_grid_5_win_rate | 0.3333333333333333 |
| decomposed_grid_6_lr_direction | 0.01 |
| decomposed_grid_6_lr_magnitude | 0.3 |
| decomposed_grid_6_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_6_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_6_success_rate | 0.3333333333333333 |
| decomposed_grid_6_win_rate | 0.3333333333333333 |
| decomposed_grid_7_lr_direction | 0.01 |
| decomposed_grid_7_lr_magnitude | 0.35 |
| decomposed_grid_7_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_grid_7_mean_final_accuracy | 0.031250000620881714 |
| decomposed_grid_7_success_rate | 0.3333333333333333 |
| decomposed_grid_7_win_rate | 0.3333333333333333 |
| decomposed_grid_8_lr_direction | 0.015 |
| decomposed_grid_8_lr_magnitude | 0.2 |
| decomposed_grid_8_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_grid_8_mean_final_accuracy | 0.027777778605620067 |
| decomposed_grid_8_success_rate | 0.3333333333333333 |
| decomposed_grid_8_win_rate | 0.3333333333333333 |
| decomposed_grid_9_lr_direction | 0.015 |
| decomposed_grid_9_lr_magnitude | 0.25 |
| decomposed_grid_9_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_grid_9_mean_final_accuracy | 0.027777778605620067 |
| decomposed_grid_9_success_rate | 0.3333333333333333 |
| decomposed_grid_9_win_rate | 0.3333333333333333 |
| decomposed_lr_direction | 0.02 |
| decomposed_lr_direction_0_mean_accuracy_gap_decomposed_minus_eggroll | 0.006944444340964158 |
| decomposed_lr_direction_0_mean_final_accuracy | 0.031250000620881714 |
| decomposed_lr_direction_0_success_rate | 0.3333333333333333 |
| decomposed_lr_direction_0_value | 0.005 |
| decomposed_lr_direction_0_win_rate | 0.3333333333333333 |
| decomposed_lr_direction_1_mean_accuracy_gap_decomposed_minus_eggroll | 0.0060763888371487456 |
| decomposed_lr_direction_1_mean_final_accuracy | 0.030381945117066305 |
| decomposed_lr_direction_1_success_rate | 0.3333333333333333 |
| decomposed_lr_direction_1_value | 0.01 |
| decomposed_lr_direction_1_win_rate | 0.3333333333333333 |
| decomposed_lr_direction_2_mean_accuracy_gap_decomposed_minus_eggroll | 0.0034722223257025084 |
| decomposed_lr_direction_2_mean_final_accuracy | 0.027777778605620067 |
| decomposed_lr_direction_2_success_rate | 0.3333333333333333 |
| decomposed_lr_direction_2_value | 0.015 |
| decomposed_lr_direction_2_win_rate | 0.3333333333333333 |
| decomposed_lr_magnitude | 0.02 |
| decomposed_lr_magnitude_0_mean_accuracy_gap_decomposed_minus_eggroll | 0.004629629664123058 |
| decomposed_lr_magnitude_0_mean_final_accuracy | 0.028935185944040615 |
| decomposed_lr_magnitude_0_success_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_0_value | 0.2 |
| decomposed_lr_magnitude_0_win_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_1_mean_accuracy_gap_decomposed_minus_eggroll | 0.005787037002543609 |
| decomposed_lr_magnitude_1_mean_final_accuracy | 0.030092593282461166 |
| decomposed_lr_magnitude_1_success_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_1_value | 0.25 |
| decomposed_lr_magnitude_1_win_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_2_mean_accuracy_gap_decomposed_minus_eggroll | 0.005787037002543609 |
| decomposed_lr_magnitude_2_mean_final_accuracy | 0.030092593282461166 |
| decomposed_lr_magnitude_2_success_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_2_value | 0.3 |
| decomposed_lr_magnitude_2_win_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_3_mean_accuracy_gap_decomposed_minus_eggroll | 0.005787037002543609 |
| decomposed_lr_magnitude_3_mean_final_accuracy | 0.030092593282461166 |
| decomposed_lr_magnitude_3_success_rate | 0.3333333333333333 |
| decomposed_lr_magnitude_3_value | 0.35 |
| decomposed_lr_magnitude_3_win_rate | 0.3333333333333333 |
| decomposed_mean_accuracy_improvement | 0.012442130129784346 |
| decomposed_mean_best_accuracy | 0.031250000620881714 |
| decomposed_mean_final_accuracy | 0.029803241447856028 |
| decomposed_mean_final_greedy_accuracy | 0.02864583410943548 |
| decomposed_mean_final_greedy_parseable_rate | 1 |
| decomposed_mean_final_parseable_rate | 0.9904514253139496 |
| decomposed_mean_finite_epoch_fraction | 0 |
| decomposed_mean_greedy_accuracy_improvement | 0.0008680555038154125 |
| decomposed_mean_greedy_parseable_rate_improvement | 0 |
| decomposed_mean_parseable_rate_improvement | -0.002604136864344279 |
| decomposed_mean_shaped_epoch_fraction | 0.16666666666666666 |
| decomposed_mean_train_accuracy | 0.022026909722222224 |
| decomposed_mean_train_parseable_rate | 0.9896918402777778 |
| decomposed_median_final_accuracy | 0.03125 |
| decomposed_success_rate | 0.3333333333333333 |
| decomposed_tangent_project_direction | 0 |
| decomposed_win_rate | 0.3333333333333333 |
| eggroll_mean_accuracy_improvement | 0.006944444961845875 |
| eggroll_mean_best_accuracy | 0.031250000620881714 |
| eggroll_mean_final_accuracy | 0.024305556279917557 |
| eggroll_mean_final_greedy_accuracy | 0.027777778605620067 |
| eggroll_mean_final_greedy_parseable_rate | 1 |
| eggroll_mean_final_parseable_rate | 0.9930555820465088 |
| eggroll_mean_finite_epoch_fraction | 0 |
| eggroll_mean_greedy_accuracy_improvement | 0 |
| eggroll_mean_greedy_parseable_rate_improvement | 0 |
| eggroll_mean_parseable_rate_improvement | 1.9868214925130207e-08 |
| eggroll_mean_shaped_epoch_fraction | 0.4166666666666667 |
| eggroll_mean_train_accuracy | 0.009114583333333334 |
| eggroll_mean_train_parseable_rate | 0.984375 |
| eggroll_median_final_accuracy | 0.02083333395421505 |
| eggroll_success_rate | 0.3333333333333333 |
| epochs | 4 |
| eval_interval | 0 |
| eval_samples | 96 |
| generations_per_prompt | 8 |
| mean_accuracy_gap_decomposed_minus_eggroll | 0.005497685167938471 |
| mean_greedy_accuracy_gap_decomposed_minus_eggroll | 0.0008680555038154125 |
| mean_greedy_improvement_gap_decomposed_minus_eggroll | 0.0008680555038154125 |
| mean_improvement_gap_decomposed_minus_eggroll | 0.005497685167938471 |
| mean_parseable_gap_decomposed_minus_eggroll | -0.002604156732559204 |
| num_decomposed_lr_directions | 3 |
| num_decomposed_lr_magnitudes | 4 |
| num_pairs | 36 |
| num_ranks | 1 |
| num_seeds | 3 |
| num_target_modules | 7 |
| population | 64 |
| prompts_per_epoch | 8 |
| rank_1_decomposed_success_rate | 0.3333333333333333 |
| rank_1_decomposed_win_rate | 0.3333333333333333 |
| rank_1_eggroll_success_rate | 0.3333333333333333 |
| rank_1_mean_accuracy_gap_decomposed_minus_eggroll | 0.005497685167938471 |
| rank_1_mean_greedy_accuracy_gap_decomposed_minus_eggroll | 0.0008680555038154125 |
| wall_clock_seconds | 18648.94031873811 |
Interpretation
Run 20260803-161316-pretrained_llm_gsm8k_scale_confirm_jean_zay-ef04761
supports the scale-calibration diagnosis, but still only weakly. The narrowed
grid again favoured decomposed EGGROLL on sampled exact-answer accuracy, with
the lower direction learning-rate region looking best. Greedy accuracy barely
moved, and the absolute GSM8K accuracy remains very low, so this is still a
scale-selection result rather than evidence that decomposed ES is already
solving the reward task. The boring alternative explanation is unchanged:
sparse exact-answer reward plus small evaluation samples can make rare lucky
generations dominate. The next useful step is to repeat the same comparison
with a stronger small model or a denser reward setting.