Spatial synaptic regularization: manuscript records and checkpoints
Companion to A human connectome reveals a synaptic spatial principle for learning while preserving memory. This repository follows the current manuscript identified in MANUSCRIPT_IDENTITY.json.
Code and reproduction guide · Panel index
Contents
Data_S1/: identical numerical source tables to the companion GitHub release, including displayed subsets and complete recorded cohorts.records/figure4/: current kernel and factorial results, original dense-editing/transfer/native-editor records, and geometry inputs.records/supplementary/S6/: complete direct FT-versus-FT+SSR transfer, 120 runs.records/supplementary/S8/: complete native-editor 250-item history evaluation of both efficacy and locality, 40 runs.records/figure5/wikirecent/: all 80 retained WikiRecent low-rank runs, four ranks and ten pairs/rank.records/figure6/andrecords/supplementary/: visual records and complete supporting cohort summaries.checkpoint_manifest.json: 14 inference checkpoints with size, tensor count, SHA-256 and current panel mapping.FILE_MANIFEST.csv: unique current paths; every listed ordinary file exists, while large tensor files use Git LFS.
The checkpoints comprise ten VOC primary tensors (two arms × five displayed seeds), two VOC replication tensors for the illustrated seed, and two CUB rank-32 tensors. They are inference states without optimizer state, not checkpoints for every experimental run. Original pretrained language-model weights are not redistributed.
Interpretation
The current main-panel source tables use five paired records wherever a paired cohort is reported: Fig. 4C/D/F and the controlled Table S6/Table S8 summaries. Raw additional observations remain available for provenance and sensitivity checks. The original Fig. 4G uses complete recipes with n=1 per setting and incomplete original seed/window metadata; it is distinct from the controlled five-pair direct FT+SSR study in Table S6. Fig. 4H uses three illustrative orders/editor at 250 edits, sampled history efficacy and immediate locality; Table S8 separately evaluates five paired orders with complete history for both metrics. Null and negative outcomes for all retained studies remain included. The archived illustrative panels and the controlled five-pair summaries are identified separately in Data_S1/figure4/COHORT_REGISTRY.md.
Usage
Download a named checkpoint with huggingface_hub.hf_hub_download, using a pinned revision for a frozen analysis. Verify its SHA-256 against checkpoint_manifest.json, then load with safetensors or the companion code's scripts/load_visual_checkpoint.py. The model architecture, dataset preprocessing and evaluation protocol must match the indexed experiment. Dataset access and upstream licenses remain with their original providers.
This is a research artifact, not a general-purpose deployed model. Exact scientific figures remain in the companion code repository. Meeting notes, editorial revisions, abandoned branches and unrelated experiments are excluded; scientific configuration and source hashes are retained.