Design

Which solid notes this experiment tests, and how.

Results

20260901-100051-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag40_jean_zay-fa7d6c8

commit: fa7d6c8

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed0
decomposed_lr_direction1
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_records13
num_methods1
num_ranks1
num_seeds1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
temperature1
top_k0
train_batch_size6
wall_clock_seconds34256.93529732688

Interpretation

Run 20260901-100051-qwen3_4b_gsm8k_decomposed_lrdir10_lrmag40_jean_zay-fa7d6c8
is a partial calibration point. It timed out before the final summary and is
therefore not directly comparable with completed grid points. Its intermediate
records are still useful for checking whether the most aggressive direction
and magnitude combination was plainly unstable.

The boring alternative explanation is that the missing final result is a
runtime/evaluation-cost issue rather than a learning failure. Keep this as a
diagnostic boundary case, not as a candidate scale unless a cheaper rerun is
explicitly needed.