Move the results-at-a-glance table directly below the opening pitch and above the caveat. Fill the completed native-grouped 10% and 20% full-MMLU values, add ratio-labelled downstream rows, sync the compact summary, and publish the latest report.
Fill the 20% MMLU-STEM 5-shot value in the top-level table, extend the labelled downstream comparison, sync the compact summary, and publish the latest progress report.
Fill the recommended 10% MMLU-STEM 5-shot value in the top-level table, label the downstream MMLU-STEM rows by pruning ratio, sync the compact summary, and publish the latest progress report.
Add top-level baseline and pruning comparison table
Add a compact results-at-a-glance table near the top of the card for baseline, 10%, and 20% native-grouped pruning. Label unavailable grouped MMLU cells explicitly as not run rather than mixing in the separate atomic-pruning sweep.
Compare importance-ranked and random pruning quality curves
Overlay the measured random 10% and 25% controls on the pruning-quality figure and update the model-card insight to foreground the advantage of HEAPr-style importance ranking.
Add random 10% pruning control and audit table labels
Record the completed random native-grouped 10% control, strengthen the importance-ranking insight at the recommended operating point, sync the progress report, and verify that README table rows have distinct labels including all GSM8K pruning ratios.
Tighten the final submission wording, label the 25% GSM8K curve rows explicitly, clarify the exploratory repacked evaluator, and state that this repository is an artifact and code package rather than a directly loadable pruned checkpoint.
Make the layer-wise score chart explicit that its y-axis is log10 importance. Explain in the card that negative tick labels denote small positive scores below one, rather than negative importance values.
Plot atomic importance distributions across sparse layers
Replace the pruning-count chart with layer-wise atomic-score box plots. Reframe the insight directly: later sparse layers contain lower-importance atomic experts under the calibration set and HEAPr-style score.
Record CRUXEval smoke validation and final time-box status
Refresh the progress report and document that CRUXEval-O CoT executable grading was smoke-tested successfully, while the paired quality comparison was stopped due to the remaining time box. Keep the public recommendation centered on the validated 10% structured-pruning point.
Pivot recommended deployment target to 10% pruning
Make 10% structured MoE pruning the recommended operating point based on the completed perplexity and GSM8K curves. Publish the 10% mask, projected 3.14B removable parameters / 9.39% full-model reduction, refreshed figures, and retain 25% as an exploratory stress point.
Add the paired GSM8K-CoT 20% structured-pruning point, revise the downstream insight to reflect increasingly steep reasoning-quality degradation, and add a reproducible seaborn curve plot. Record the newly started CRUXEval-O CoT evaluation.
Publish the completed paired GSM8K-CoT 10% pruning point and note the preliminary task-specific threshold pattern relative to the 25% result while the 20% point runs.
Publish the completed same-size random-control loss result and document that HEAPr-ranked structured pruning preserves perplexity materially better than random selection under the same minimum-retention rule.
Add the exact structured-pruning parameter calculation: 7.85B removable routed-MoE parameters, 23.48% of the full model, with a projected 25.59B compact checkpoint and 14.625 GiB BF16 weight-storage reduction.
Lead with the deployment-oriented 25% structured MoE parameter-removal claim, move atomic/group terminology into the method explanation, and clarify that compact checkpoint and kernel benchmarks remain future work.
Publish Laguna Martini grouped-pruning model card and reproducibility artifacts
Add the HEAPr-style grouped atomic-expert pruning submission: model card, curated runnable source snapshot, compact results summary, selected reports, and 25% native-group reproducibility artifacts. This intentionally excludes model weights and does not claim kernel speedup yet.
