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/: the same source files as the current companion GitHub release.records/figure4/: kernel, factorial, complete-recipe transfer, AlphaEdit, and geometry records for the reported comparisons.records/figure5/wikirecent/: WikiRecent low-rank runs at ranks 8, 16 and 32, ten pairs per rank.records/figure6/andrecords/supplementary/: visual records and complete supporting cohort summaries.checkpoint_manifest.json: 14 visual inference checkpoints and 40 final Fig. 5 follow-up LoRA adapters, with SHA-256 and panel mapping.FILE_MANIFEST.csv: file sizes and SHA-256 for the current repository paths; tensor files use large-file storage.
The visual checkpoints comprise ten VOC primary tensors, two VOC replication tensors, and two CUB rank-32 tensors. The 40 WikiRecent adapters cover both arms of ten paired orders at each of ranks 8 and 32 after 100 edits. They are final inference states without optimizer state; earlier trajectory snapshots are not uploaded. Original pretrained language-model weights are not redistributed.
Interpretation
Figure 4B/E report five paired seeds; C/D/F use ten paired orders per comparison. Figure 4G is one complete-recipe pair per setting, and H shows ten paired AlphaEdit orders with both efficacy and locality evaluated on all 250 historical items. Figure 5A/B/E/F use ten paired rank-32 follow-up orders; C shows ranks 8, 16 and 32 from the eight-target sweep, while D shows immediate and history locality at ranks 8, 16 and 32 under 32-target rehearsal. Seed lists are 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.
The records/ tree preserves the earlier archived evaluations. Data_S1/ is the source data for the reported experiments.