Design

Protocol control for Qwen2.5-0.5B GSM8K evaluation. The run compares the unpatched base model with rank-1 EGGROLL and decomposed EGGROLL patched models under zero updates, using the same 8-shot CoT prompt/parser as the baseline sanity eval. The purpose is to verify that the ES adapter wrappers do not change behaviour before optimisation.

Results

20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3

commit: 1a327b3

metricvalue
base_accuracy0.38438212871551514
base_num_correct507
base_num_parseable1319
base_parseable_rate1
decomposed_patched_accuracy0.38438212871551514
decomposed_patched_accuracy_gap_vs_base0
decomposed_patched_num_correct507
decomposed_patched_num_parseable1319
decomposed_patched_parseable_gap_vs_base0
decomposed_patched_parseable_rate1
eggroll_patched_accuracy0.38438212871551514
eggroll_patched_accuracy_gap_vs_base0
eggroll_patched_num_correct507
eggroll_patched_num_parseable1319
eggroll_patched_parseable_gap_vs_base0
eggroll_patched_parseable_rate1
eval_batch_size128
eval_samples1319
max_new_tokens512
num_target_modules7
rank1
wall_clock_seconds1003.3977364134043

Interpretation

Run 20260809-005632-qwen05_gsm8k_es_protocol_control_jean_zay-1a327b3 passes the wrapper-agreement check: base, EGGROLL-patched, and decomposed-patched evaluations are identical on accuracy and parseability, with zero gaps versus base for both patched variants. This means the earlier very low Qwen smoke accuracies were not caused by merely attaching ES adapters; future Qwen reward comparisons can use the patched harness, but should still report the protocol distinction from the no-ES baseline eval.