2026-08-09
- 00:58 submitted 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3 (job 766494 on jean-zay, commit 1a327b3)
- 00:58 submitted fixed Qwen GSM8K ES protocol control 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3 on Jean Zay as job 766494 to rerun the no-update wrapper agreement check after batching generation to avoid OOM
- 04:59 fetched 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3
- 04:59 reported 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3 in notes/experiments/qwen05_gsm8k_es_protocol_control_jean_zay.md
- 04:59 fetched and reported Qwen GSM8K ES protocol control 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3; base, EGGROLL-patched, and decomposed-patched no-update evaluations matched exactly, so the adapter wrappers are not changing behaviour before optimisation
- 16:32 prepared a Qwen GSM8K destructive-update diagnostic to check whether ES updates damage train/eval behaviour by logging per-epoch train accuracy, eval accuracy, prompt-logit KL drift, reward variance, and update norms before running another broad sweep
- 16:53 fixed the Qwen destructive-update diagnostic Jean Zay bootstrap quoting after the first submit attempt failed before scheduler submission while warming the GSM8K dataset cache
- 16:55 submitted 20260809-165410-qwen05_gsm8k_destructive_update_diagnostic_jean_zay-4b5ea4f (job 789596 on jean-zay, commit 4b5ea4f)
- 16:55 submitted Qwen GSM8K destructive-update diagnostic 20260809-165410-qwen05_gsm8k_destructive_update_diagnostic_jean_zay-4b5ea4f on Jean Zay as job 789596 to test whether reward-driven ES updates damage train/eval behaviour via per-epoch accuracy, KL drift, reward variance, and update norms
- 18:57 fetched 20260809-165410-qwen05_gsm8k_destructive_update_diagnostic_jean_zay-4b5ea4f
- 18:57 reported 20260809-165410-qwen05_gsm8k_destructive_update_diagnostic_jean_zay-4b5ea4f in notes/experiments/qwen05_gsm8k_destructive_update_diagnostic_jean_zay.md
- 18:58 fetched, reported, and interpreted Qwen GSM8K destructive-update diagnostic 20260809-165410-qwen05_gsm8k_destructive_update_diagnostic_jean_zay-4b5ea4f; the run supports the destructive-update suspicion because exact reward stayed zero while parseability shaping drove KL drift and eval accuracy fell