PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (1401-1600/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (1201-1400/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (1001-1200/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (801-1000/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (601-800/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (401-600/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (201-400/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish FM c6879962 training artifacts (1-200/1676)
PF FM 500k campaign: experts, joint training and fine-tunes; latest and best checkpoints, complete histories, TensorBoard, logs, configs, manifests, original-step plots and integrity metadata.
Publish CMS Attention and HEPTv2 320k training from e2c8e65c
Final 320k checkpoints, full histories and resumed TensorBoard streams, all rank logs, configs, manifests, final plots, paired comparison metrics and SHA256 checksums.
Move legends directly beneath the event displays and size each canvas to its legend. Replace center-dot text separators throughout event figures and the gallery. Keep the saved event data, colors and physical scales unchanged.
Replace reconstruction diagram with CLD event flow and reusable detector views
Render one event per detector family with identical component layers, white and transparent exports, vector SVGs, and extracted scene provenance. Show reconstructed Pandora particles for CLD, CLIC and MAIA; explicitly label truth proxies for IDEA and ColliderML. Highlight track/cluster and raw-hit MLPF routes.
Refer dataset guidance to docs and record v3.2.0 training selections
Use independent loss axes for pileup and no-pileup groups. Retain full-history panels and add zoom panels for steps 60000 to 100000, preserving the existing data and smoothing.
Plot physical_nopu and physical_pu group loss in separate CMS panels. Retain 2000-step half-life smoothing for training and unsmoothed validation. Explain group aggregation and dataset mixture proportions.
Use TensorBoard loss series and smooth training curves
Use step/loss_Total_unweighted from train and validation TensorBoard streams. Smooth training with an exponential moving average of 2000 optimizer-step half-life; show validation unsmoothed. Include raw values, smoothed values and source provenance.
Simplify model card and show loss and physics evolution curves
Show training and validation loss for CMS, CLIC and CLD, plus six physics metrics versus step for CLIC and CLD. Remove CMS physics results from the model card and shorten the explanatory text. Include curve values and source provenance.
Clarify GPU memory footprint and add ROCm utilization averages
Distinguish tensor allocation, allocator reservation and device-wide VRAM occupancy. GPU utilization averages devices 1–7 for hit runs, retains intermittent zeros, excludes logging-biased device 0, and marks PF utilization unavailable. Include full snapshot audit data.
Update model card with v3.2.0 model links and physics comparisons
Link all ten current CMS, CLIC and CLD checkpoints. Add separate physics comparison charts, exact source data, training resource table and evaluation caveats. Preserve papers and dataset references.
Add clic/clusters/v3.2.0: attention and HEPTv2 final 50k checkpoints and training evidence
Training revision 6b80b983477321cf708e5043fb1d2ec15dd398bc. Preserve original checkpoints, complete available histories, all logs, configs, TensorBoard streams and final plots. Add release metadata and comparison summaries. No deletions or code tags.
Add clic/hits/v3.2.0: attention and HEPTv2 final 50k checkpoints and training evidence
Training revision 6b80b983477321cf708e5043fb1d2ec15dd398bc. Preserve original checkpoints, complete available histories, all logs, configs, TensorBoard streams and final plots. Add release metadata and comparison summaries. No deletions or code tags.
Add cld/clusters/v3.2.0: attention and HEPTv2 final 50k checkpoints and training evidence
Training revision 6b80b983477321cf708e5043fb1d2ec15dd398bc. Preserve original checkpoints, complete available histories, all logs, configs, TensorBoard streams and final plots. Add release metadata and comparison summaries. No deletions or code tags.
Add cld/hits/v3.2.0: attention and HEPTv2 final 50k checkpoints and training evidence
Training revision 6b80b983477321cf708e5043fb1d2ec15dd398bc. Preserve original checkpoints, complete available histories, all logs, configs, TensorBoard streams and final plots. Add release metadata and comparison summaries. No deletions or code tags.
22 days ago
Add CMS v3.2.0 Attention and HEPTv2 100k checkpoints
Publish both completed October 27 CMS backbone runs with final checkpoints, complete metric histories, TensorBoard streams, configurations, manifests, rank logs and final plots. Include common-event reference comparison metrics. Training dataset version is 2.8.0.