2026-08-02

• 00:30 fetched 20260801-175851-pretrained_llm_gsm8k_reward_sweep_jean_zay-993bd00
• 00:30 reported 20260801-175851-pretrained_llm_gsm8k_reward_sweep_jean_zay-993bd00 in notes/experiments/pretrained_llm_gsm8k_reward_sweep_jean_zay.md
• 00:30 fetched and reported GSM8K reward sweep 20260801-175851-pretrained_llm_gsm8k_reward_sweep_jean_zay-993bd00; both methods stayed finite but exact-reward accuracy stayed very low, with no decomposed advantage over plain EGGROLL
• 11:40 wired GSM8K reward runner to repeat prompts by generations_per_prompt and pass grouped fitness size through plain and decomposed EGGROLL for a G=8, noise_reuse=4 retry
• 11:46 submitted 20260802-114517-pretrained_llm_gsm8k_grouped_reward_sweep_jean_zay-48f4f83 (job 533000 on jean-zay, commit 48f4f83)
• 11:47 submitted grouped GSM8K reward sweep 20260802-114517-pretrained_llm_gsm8k_grouped_reward_sweep_jean_zay-48f4f83 on Jean Zay as job 533000 with G=8, population=64, noise_reuse=4 after fixing bootstrap quoting