spikeforge

spikeforge model hub

17 catalog entries -- 7 with trained weights, 17 with a verified source and license.

This page is generated from the catalog, not a live view of what is downloaded or cached.

7 entries carry trained weights -- checkpoints this project trained itself, shipped with spikeforge-hub, each showing what it scores on the complete held-out test split. They are reference configurations with stock hyperparameters, not tuned attempts at state of the art, and every one names the command that reproduces it. The remaining entries are bundled: this project's own NIR graph presets -- structure, with freshly-initialised weights. No third-party weights are redistributed. See CURATION.md for the verification policy, and open a pull request there to propose a real, checked entry.
Name / idWeightsFrameworkKind SourceLicenseTraining data Notes
mnist / fc_legacy (94.29%)
reference/mnist-fc-legacy		 trained weights
94.29% on 10000 held-out mnist samples
download weights		 snntorch state_dict reference
fc_legacy		 BSD-3-Clause verified source mnist
unverified-candidate
The MNIST database of handwritten digits, Yann LeCun, Corinna Cortes, and Christopher J.C. Burges.		 Trained by this project on the full mnist training split: 3 epochs, 25 time steps, seed 0, CPU. Scores 94.29% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only mnist-fc-legacy. The default topology and the README quickstart's dataset.
mnist / fc_small (92.16%)
reference/mnist-fc-small		 trained weights
92.16% on 10000 held-out mnist samples
download weights		 snntorch state_dict reference
fc_small		 BSD-3-Clause verified source mnist
unverified-candidate
The MNIST database of handwritten digits, Yann LeCun, Corinna Cortes, and Christopher J.C. Burges.		 Trained by this project on the full mnist training split: 3 epochs, 25 time steps, seed 0, CPU. Scores 92.16% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only mnist-fc-small. The smaller fully-connected preset, for edge-sized budgets.
mnist / conv_net (97.13%)
reference/mnist-conv-net		 trained weights
97.13% on 10000 held-out mnist samples
download weights		 snntorch state_dict reference
conv_net		 BSD-3-Clause verified source mnist
unverified-candidate
The MNIST database of handwritten digits, Yann LeCun, Corinna Cortes, and Christopher J.C. Burges.		 Trained by this project on the full mnist training split: 2 epochs, 25 time steps, seed 0, CPU. Scores 97.13% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only mnist-conv-net. Convolutional LIF; fewer epochs because each one costs more.
mnist / recurrent_net (92.46%)
reference/mnist-recurrent-net		 trained weights
92.46% on 10000 held-out mnist samples
download weights		 snntorch state_dict reference
recurrent_net		 BSD-3-Clause verified source mnist
unverified-candidate
The MNIST database of handwritten digits, Yann LeCun, Corinna Cortes, and Christopher J.C. Burges.		 Trained by this project on the full mnist training split: 3 epochs, 25 time steps, seed 0, CPU. Scores 92.46% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only mnist-recurrent-net. Recurrent LIF with a one-step delayed feedback edge.
fashion / fc_legacy (75.55%)
reference/fashion-fc-legacy		 trained weights
75.55% on 10000 held-out fashion samples
download weights		 snntorch state_dict reference
fc_legacy		 BSD-3-Clause verified source fashion
MIT
Fashion-MNIST, Copyright (c) 2017 Zalando SE (https://tech.zalando.com), MIT licensed.		 Trained by this project on the full fashion training split: 3 epochs, 25 time steps, seed 0, CPU. Scores 75.55% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only fashion-fc-legacy. Fashion-MNIST: same shape, a harder ten-class problem.
kmnist / fc_legacy (70.79%)
reference/kmnist-fc-legacy		 trained weights
70.79% on 10000 held-out kmnist samples
download weights		 snntorch state_dict reference
fc_legacy		 BSD-3-Clause verified source kmnist
CC-BY-SA-4.0
"KMNIST Dataset" (created by CODH), adapted from "Kuzushiji Dataset" (created by NIJL and others), doi:10.20676/00000341		 Trained by this project on the full kmnist training split: 3 epochs, 25 time steps, seed 0, CPU. Scores 70.79% on all 10000 held-out samples. A reference configuration with stock hyperparameters, not a tuned attempt at state of the art. Reproduce with: python scripts/train_reference_models.py --only kmnist-fc-legacy. Kuzushiji-MNIST: cursive Japanese characters.
Legacy FC LIF (fc_legacy)
nir/fc_legacy		 structure only nir nir_graph bundled
fc_legacy		 BSD-3-Clause verified source not trained on a dataset Shipped fc_legacy preset rendered as a NIR graph. Bundled with the project; no third-party weights are redistributed.
Small FC LIF (fc_small)
nir/fc_small		 structure only nir nir_graph bundled
fc_small		 BSD-3-Clause verified source not trained on a dataset Shipped fc_small preset rendered as a NIR graph. Bundled with the project; no third-party weights are redistributed.
Convolutional LIF (conv_net)
nir/conv_net		 structure only nir nir_graph bundled
conv_net		 BSD-3-Clause verified source not trained on a dataset Shipped conv_net preset rendered as a NIR graph. Bundled with the project; no third-party weights are redistributed.
Recurrent LIF (recurrent_net)
nir/recurrent_net		 structure only nir nir_graph bundled
recurrent_net		 BSD-3-Clause verified source not trained on a dataset Shipped recurrent_net preset rendered as a NIR graph. Bundled with the project; no third-party weights are redistributed.
snnTorch FC reference
snntorch/fc_reference		 structure only snntorch nir_graph bundled
fc_small		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped fc_small preset, so a snnTorch-authored model can be compared against a known-good structure.
snnTorch recurrent reference
snntorch/recurrent_reference		 structure only snntorch nir_graph bundled
recurrent_net		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped recurrent_net preset.
SpikingJelly conv reference
spikingjelly/conv_reference		 structure only spikingjelly nir_graph bundled
conv_net		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped conv_net preset.
SpikingJelly FC reference
spikingjelly/fc_reference		 structure only spikingjelly nir_graph bundled
fc_small		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped fc_small preset.
Norse conv reference
norse/conv_reference		 structure only norse nir_graph bundled
conv_net		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped conv_net preset, used as the first simulator-backend reference target.
Lava/Loihi recurrent reference
lava/recurrent_reference		 structure only lava nir_graph bundled
recurrent_net		 BSD-3-Clause verified source not trained on a dataset Ecosystem provenance only: the artifact is the bundled NIR graph derived from the shipped recurrent_net preset, used as a declared hardware-path reference.
rsi-search-policy-w1 (transfer-verified search policy)
reference/rsi-search-policy-w1		 trained weights
100.0% on 3 held-out rsi-shadow-w1-synthetic samples
download weights ยท manifest		 snntorch state_dict reference
custom: rsi-search 12-feat linear/LIF/readout (sidecar spec)		 BSD-3-Clause verified source rsi-shadow-w1-synthetic
BSD-3-Clause
Rows generated by Capsize-Games/rsi-programmer (rsi_programmer/probe.py) and frozen in fixtures/rsi_shadow_tasks.json; no third-party data.		 SNN search-control policy: linear(12,h)/leaky(beta 0.9)/readout, 12 drive steps, seed 0, 50 pairwise-margin epochs on W1 traces + same-op contrastive pairs. Ranked search beats the plain enumerator on 3/3 held-out novel rules; beam 9/24 vs ranked enumeration. Reproduce with rsi-programmer/scripts/train.py --seed 0 --epochs 50 (Capsize-Games/rsi-programmer; see rsi-search-policy-w1.manifest.json). Weights + manifest + LICENSE ship together; the LICENSE file governs the weights. Built and trained through Spikeforge TopologySpec/build_module. test_accuracy here means task-level solve rate: 3/3 held-out novel rules solved by SNN-guided ranked search (158/122/957 tried vs 939/939/6807 enumerator), not sample classification accuracy.