Design

Which solid notes this experiment tests, and how.

Results

20260901-095300-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag05_jean_zay-5a44d6a

commit: 5a44d6a

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed0
decomposed_lr_direction0.5
decomposed_lr_magnitude0.5
decomposed_sigma_direction0.001
decomposed_sigma_magnitude0.001
decomposed_tangent_project_direction0
eggroll_lr1
eggroll_sigma0.001
epochs15
eval_batch_size8
eval_prompts128
generations_per_prompt6
max_new_tokens1024
method_decomposed_enabled1
method_plain_enabled0
noise_reuse4
num_center_eval_records3
num_completed_runs0
num_epoch_records15
num_methods1
num_ranks1
num_seeds1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
temperature1
top_k0
train_batch_size6
wall_clock_seconds35373.568909072084

Interpretation

Run 20260901-095300-qwen3_4b_gsm8k_decomposed_lrdir05_lrmag05_jean_zay-5a44d6a
is a partial calibration point. It reached the epoch loop and flushed center
checks, but it did not complete the final summary before the time limit. That
makes it useful for trend inspection, not for declaring this scale better or
worse than completed settings.

The boring alternative explanation is again runtime pressure from the
evaluation protocol. This setting should be compared at matched intermediate
checkpoints or rerun with cheaper evaluation before it is used to choose the
decomposed method scale.