Design

This page merges the matrix-only experiments that validated the ES plumbing before any LLM benchmark. These runs test controlled low-rank matrix learning on Jean Zay. They are implementation and geometry checks, not evidence about language-task learning.

The sequence was:

  1. plain EGGROLL matrix-learning smoke;
  2. single-seed plain-vs-decomposed comparison;
  3. seed/rank robustness sweep with full, direction-only, and magnitude-only decomposed ablations.

Results

20260731-115158-eggroll_matrix_learning_jean_zay-5efd8d0

commit: 5efd8d0

metricvalue
initial_loss2.9565727710723877
final_loss0.203643798828125
best_loss0.203643798828125
finite_epoch_fraction1
success1

20260731-140705-eggroll_vs_decomposed_matrix_learning_jean_zay-05ce937

commit: 05ce937

metricvalue
eggroll_final_loss0.203643798828125
decomposed_final_loss0.09557401388883591
loss_gap_decomposed_minus_eggroll-0.10806978493928909
eggroll_success1
decomposed_success1
decomposed_wins1

20260731-162711-matrix_robustness_sweep_jean_zay-04fee3e

commit: 04fee3e

metricvalue
num_seeds10
num_ranks4
num_records160
eggroll_success_rate0.65
dora_full_success_rate1
dora_direction_only_success_rate0
dora_magnitude_only_success_rate0
eggroll_mean_final_loss1.91025531408377
dora_full_mean_final_loss0.39386328402906656
dora_direction_only_mean_final_loss4.286366328597069
dora_magnitude_only_mean_final_loss4.3086372256278995
dora_full_win_rate_vs_eggroll0.8
dora_full_mean_loss_gap_vs_eggroll-1.5163920300547034

Interpretation

The matrix sequence did what it was meant to do. Plain EGGROLL learned the toy low-rank matrix objective, which made it credible as the baseline implementation. The decomposed DoRA-style variant then beat plain EGGROLL in the single-seed comparison and, more importantly, in most paired comparisons in the seed/rank sweep.

The ablations are important: direction-only and magnitude-only decomposed variants both failed on the sweep, while the full decomposed method succeeded in every tested condition. That supports the specific magnitude-plus-direction story rather than a generic “more parameters” story.

This evidence is still synthetic and matrix-only. It justifies moving to LLM smoke tests and then to a real reward benchmark; it is not a final project claim.