Design

Which solid notes this experiment tests, and how.

Results

20260901-095021-qwen3_4b_gsm8k_decomposed_lrdir025_lrmag20_jean_zay-fb73f8f

commit: fb73f8f

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed0
decomposed_lr_direction0.25
decomposed_lr_magnitude2
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_seconds33716.349802547134

Interpretation

Run 20260901-095021-qwen3_4b_gsm8k_decomposed_lrdir025_lrmag20_jean_zay-fb73f8f
is a partial calibration point. It timed out before producing a completed run
summary, so it should not be treated as a final score for this learning-rate
setting. The flushed epoch and center-eval records are still useful for
diagnosing whether this region of the grid is promising or too slow under the
current evaluation budget.

The boring alternative explanation is runtime rather than learning dynamics:
this configuration may simply need a cheaper final evaluation path to become
comparable with the completed grid points. Use it as partial evidence only
until it is rerun or compared through matched intermediate checkpoints.