2026-08-05

• 08:57 fetched 20260804-134811-qwen05_gsm8k_baseline_eval_jean_zay-889d9d3
• 08:58 reported 20260804-134811-qwen05_gsm8k_baseline_eval_jean_zay-889d9d3 in notes/experiments/qwen05_gsm8k_baseline_eval_jean_zay.md
• 08:58 fetched and reported Qwen2.5-0.5B GSM8K baseline sanity eval 20260804-134811-qwen05_gsm8k_baseline_eval_jean_zay-889d9d3 so the low Qwen smoke accuracy can be compared against a no-ES protocol check
• 09:47 implemented a Qwen GSM8K ES protocol control so the reward harness can be checked with the same 8-shot CoT prompt and final-answer parser as the no-ES baseline before spending on new ES comparisons
• 10:06 fixed Qwen GSM8K ES protocol-control Jean Zay bootstrap quoting after the first explicit submit attempt failed before scheduler submission while warming the GSM8K dataset cache
• 10:08 submitted 20260805-100718-qwen05_gsm8k_es_protocol_control_jean_zay-1dd4c03 (job 597144 on jean-zay, commit 1dd4c03)
• 10:09 submitted Qwen GSM8K ES protocol control 20260805-100718-qwen05_gsm8k_es_protocol_control_jean_zay-1dd4c03 on Jean Zay as job 597144 to check base, EGGROLL-patched, and decomposed-patched no-update 8-shot CoT eval agreement
• 12:20 fetched 20260805-100718-qwen05_gsm8k_es_protocol_control_jean_zay-1dd4c03
• 12:20 fetched Qwen GSM8K ES protocol control run 20260805-100718-qwen05_gsm8k_es_protocol_control_jean_zay-1dd4c03 after Jean Zay job 597144 failed with CUDA OOM before producing metrics; xp report could not run because results/metrics.json is absent
• 15:10 changed the Qwen GSM8K ES protocol control to evaluate in batches of 128 after the full-test-set batch OOMed on H100; the failure was generation batch size, not model size