Design
Which solid notes this experiment tests, and how.
Results
20260802-114517-pretrained_llm_gsm8k_grouped_reward_sweep_jean_zay-48f4f83
commit: 48f4f83
| metric | value |
|---|---|
| both_success_rate | 0.3333333333333333 |
| decomposed_mean_accuracy_improvement | 0.008680556124697128 |
| decomposed_mean_best_accuracy | 0.03645833395421505 |
| decomposed_mean_final_accuracy | 0.026041667442768812 |
| decomposed_mean_final_greedy_accuracy | 0.027777778605620067 |
| decomposed_mean_final_greedy_parseable_rate | 0.9947916865348816 |
| decomposed_mean_final_parseable_rate | 0.9774305621782938 |
| decomposed_mean_finite_epoch_fraction | 1 |
| decomposed_mean_greedy_accuracy_improvement | 0 |
| decomposed_mean_greedy_parseable_rate_improvement | -0.0017361044883728027 |
| decomposed_mean_parseable_rate_improvement | 0.005208313465118408 |
| decomposed_mean_shaped_epoch_fraction | 0.375 |
| decomposed_mean_train_accuracy | 0.016276041666666668 |
| decomposed_mean_train_parseable_rate | 0.9700520833333334 |
| decomposed_median_final_accuracy | 0.02083333395421505 |
| decomposed_success_rate | 0.3333333333333333 |
| decomposed_win_rate | 0 |
| eggroll_mean_accuracy_improvement | 0.012152778295179209 |
| eggroll_mean_best_accuracy | 0.038194445272286735 |
| eggroll_mean_final_accuracy | 0.029513889613250893 |
| eggroll_mean_final_greedy_accuracy | 0.027777778605620067 |
| eggroll_mean_final_greedy_parseable_rate | 0.9965277910232544 |
| eggroll_mean_final_parseable_rate | 0.9809027910232544 |
| eggroll_mean_finite_epoch_fraction | 1 |
| eggroll_mean_greedy_accuracy_improvement | 0 |
| eggroll_mean_greedy_parseable_rate_improvement | 0 |
| eggroll_mean_parseable_rate_improvement | 0.00868054231007894 |
| eggroll_mean_shaped_epoch_fraction | 0.4583333333333333 |
| eggroll_mean_train_accuracy | 0.013997395833333334 |
| eggroll_mean_train_parseable_rate | 0.9710286458333334 |
| eggroll_median_final_accuracy | 0.026041666977107525 |
| eggroll_success_rate | 0.3333333333333333 |
| epochs | 8 |
| eval_samples | 96 |
| generations_per_prompt | 8 |
| mean_accuracy_gap_decomposed_minus_eggroll | -0.003472222170482079 |
| mean_greedy_accuracy_gap_decomposed_minus_eggroll | 0 |
| mean_greedy_improvement_gap_decomposed_minus_eggroll | 0 |
| mean_improvement_gap_decomposed_minus_eggroll | -0.003472222170482079 |
| mean_parseable_gap_decomposed_minus_eggroll | -0.0034722288449605307 |
| num_pairs | 6 |
| num_ranks | 2 |
| 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 |
| rank_1_eggroll_success_rate | 0.3333333333333333 |
| rank_1_mean_accuracy_gap_decomposed_minus_eggroll | -0.0034722223257025084 |
| rank_1_mean_greedy_accuracy_gap_decomposed_minus_eggroll | 0 |
| rank_2_decomposed_success_rate | 0.3333333333333333 |
| rank_2_decomposed_win_rate | 0 |
| rank_2_eggroll_success_rate | 0.3333333333333333 |
| rank_2_mean_accuracy_gap_decomposed_minus_eggroll | -0.00347222201526165 |
| rank_2_mean_greedy_accuracy_gap_decomposed_minus_eggroll | 0 |
Interpretation
Run 20260802-114517-pretrained_llm_gsm8k_grouped_reward_sweep_jean_zay-48f4f83 completed cleanly and confirms the grouped GSM8K reward path is stable. It does not support a decomposed advantage: plain EGGROLL was slightly ahead on sampled exact-answer accuracy, greedy accuracy was tied, and both methods remained in the same very-low-reward regime. The grouped sampling/reuse change helped the measurement plumbing more than the learning signal.