Design

Which solid notes this experiment tests, and how.

Results

20260901-095158-qwen3_4b_gsm8k_decomposed_lrdir025_lrmag40_jean_zay-ab9bbb8

commit: ab9bbb8

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed0
decomposed_lr_direction0.25
decomposed_lr_magnitude4
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_records14
num_methods1
num_ranks1
num_seeds1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
temperature1
top_k0
train_batch_size6
wall_clock_seconds34026.89495655312

Interpretation

Run 20260901-095158-qwen3_4b_gsm8k_decomposed_lrdir025_lrmag40_jean_zay-ab9bbb8
is a partial calibration point. It timed out before the final center-model
evaluation and completed-run summary, so the setting should not be ranked
against completed runs on final accuracy. The partial records are enough to
check whether the high magnitude step showed early instability under exact
reward, but not enough for a selection decision by themselves.

The boring alternative explanation is that the fixed batch and expensive
evaluation schedule, rather than this scale choice, caused the missing final
summary. Keep this point in the diagnostic set, but prefer completed runs when
choosing the next repeated-seed test.