SPIRA v3: cortex representation repair — direct input pathway, multi-timescale trace banks, structured trigram inputs; controlled A/B (SNN alone 0.40% -> 2.07% in-vocab held-out vs bigram 4.94% at matched vocab 4096); paper, README and docs updated to third iteration
6923c51 verified Artifacts
| File | What |
|---|---|
spira_model.npz (72 MB) |
full trained v2 checkpoint: EPT stats, reservoir + slow area (W_out, W_out_G, homeostatic state incl. β), fusion calibration, hippocampus (exact derivable-slot round-trip) — resume with SPIRA.load() |
spira_model.npz.raster.npz |
spike raster recorded over the last 1500 ticks of training (feeds the Brain tab) |
metrics.json |
full v2 evaluation: coverage curve, compression, prediction + hierarchy ablation, one-shot, replay consolidation experiment, correction (two load regimes), generation, efficiency |
spira_model_v3_ab.npz (45.9 MB) |
v3 A/B checkpoint — vocab 4096, trained under the controlled A/B protocol (200k tokens × 2 epochs, AM 50k contexts) for the cortex-representation-repair experiment; it is not a replacement for the main 16k-vocab v2 checkpoint spira_model.npz above. Load with SPIRA.load() using the v3 code in ../spira/. |
metrics_v3_ab.json |
v3 A/B results: protocol, both arms' full metrics, the matched-vocab bigram reference (4.94%), and the headline comparison (SNN alone, in-vocab, held-out: v2 0.40% → v3 2.07%) — see ../../V3_REPORT.md |
chunks.jsonl |
training-chunk log (historical) |
web_*.png |
browser-demo screenshots |
spira_deploy.json (29 MB, browser-inference export, format v2) is not duplicated
here — it lives at web/public/spira_deploy.json, where the demo fetches it.
Regenerate it with:
python3 -m spira.export_deploy --model artifacts/spira_model.npz \
--out ../web/public/spira_deploy.json --am-cap 100000 --am-min-count 1
Figures are generated by scripts/make_figures.py and stored with the paper in
../paper/figures/ (incl. fig_vocab.pdf and fig_replay.pdf).