Design

Which solid notes this experiment tests, and how.

Results

20260826-093813-qwen3_4b_gsm8k_fixed_batch_reward_signal_negative_jean_zay-620a35d

commit: 620a35d

metricvalue
center_eval_do_sample0
center_eval_every5
center_eval_temperature1
center_eval_top_k0
completed1
condition_0_best_center_eval_accuracy0.3828125
condition_0_center_eval_accuracy_change0.0234375
condition_0_center_train_accuracy_change0.0476190485060215
condition_0_condition_index0
condition_0_final_center_eval_accuracy0.1484375
condition_0_final_center_eval_parseable_rate1
condition_0_final_center_train_accuracy0.095238097012043
condition_0_final_center_train_parseable_rate1
condition_0_initial_center_eval_accuracy0.125
condition_0_initial_center_eval_parseable_rate1
condition_0_initial_center_train_accuracy0.0476190485060215
condition_0_initial_center_train_parseable_rate1
condition_0_lr-1
condition_0_mean_fitness_std0.9027024000883103
condition_0_mean_perturbed_train_accuracy0.09841270502656699
condition_0_mean_perturbed_train_parseable_rate1
condition_0_mean_raw_score_std0.28806543424725534
condition_0_nonzero_reward_epoch_fraction1
condition_0_num_layers253
condition_0_rank1
condition_0_seed0
condition_0_shaped_epoch_fraction0
eggroll_sigma0.001
epochs20
eval_batch_size8
eval_prompts128
generations_per_prompt6
max_new_tokens1024
noise_reuse4
num_center_eval_records5
num_completed_runs1
num_epoch_records20
num_lr_conditions1
num_target_module_patterns8
parseability_shaping0
population126
prompts_per_epoch21
rank1
seed0
temperature1
top_k0
train_batch_size6
wall_clock_seconds22194.486616202048

Interpretation

Run 20260826-093813-qwen3_4b_gsm8k_fixed_batch_reward_signal_negative_jean_zay-620a35d
completed the missing negative-learning-rate control from the fixed-batch
reward-signal diagnostic. The negative update direction did not reproduce the
large sustained center-model gain seen in the earlier positive-learning-rate
condition from
20260824-183902-qwen3_4b_gsm8k_fixed_batch_reward_signal_jean_zay-39c4380.

This supports the boring-but-important interpretation that the positive control
was seeing a usable exact-answer reward signal on that fixed batch, not merely a
generic effect of taking large EGGROLL steps or of parseability shaping.
Parseability was already saturated and shaping was disabled, so those are weak
alternative explanations here. The remaining caveat is that this is still one
fixed batch and one seed; it is a reward-signal diagnostic, not a robust
decomposed-versus-plain comparison.

Next, use this as permission to move from “is there any scalar reward signal?”
to a small repeated-seed or easier-curriculum test before interpreting method
differences.