Design

Which solid notes this experiment tests, and how.

Results

20260901-095835-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag10_jean_zay-f89db94

commit: f89db94

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed0
decomposed_lr_direction1
decomposed_lr_magnitude1
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_seconds35739.84462727909

Interpretation

Run 20260901-095835-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag10_jean_zay-f89db94
is a partial calibration point. It timed out before producing the completed-run
summary, so it should be treated as an intermediate trace only. The setting may
still be informative when compared at matched center checkpoints across the
grid.

The boring alternative explanation is wall-clock budget rather than a failed
learning rate. Do not promote this setting without a completed rerun or a
clear intermediate-checkpoint advantage over completed candidates.