SNN-4-ALL MICRO 2026 AE Checkpoints
This repository contains checksum-verified checkpoints prepared
for the MICRO 2026 Artifact Evaluation package of:
SNN-4-All: Breaking the Edge Fine-Tuning Barrier with Heterogeneous Sparsity
Exploitation
Source artifact
- GitHub repository:
GDHNES/SNN-4-ALL-AE-release
- Planned evaluated release:
micro2026-ae-v1.1
- The exact evaluated source commit is recorded by the tagged GitHub release
and the corresponding archival record.
Included checkpoints
- Server SpikingVGG-11 baseline, coupled-protocol epoch 30
- Server SpikingVGG-11 TwiST (0.7-2), coupled-protocol epoch 24
- Orin Spikformer baseline historical best checkpoint
- Orin Spikformer TwiST historical best checkpoint
The Server SpikingVGG-11 baseline and TwiST checkpoints were produced by full
training under the recovered coupled historical protocol. Their independently
evaluated accuracies were 58.1% and 59.5%, versus 59.4% and 60.2% reported in
the paper; both satisfy the pre-declared +/-3 percentage-point tolerance.
The Orin Spikformer checkpoints support checksum-verified, complete-test and
resource replay. This is checkpoint-based replay evidence, not a new
full-training reproduction.
Integrity and license
Verify every checkpoint against CHECKSUMS.sha256 before use. Exact provenance
and the redistribution license are documented in
CHECKPOINT_LICENSES.md and checkpoint_manifest.json.
The public SpikeGPT-216M checkpoint is not duplicated here. The source artifact
downloads the upstream Apache-2.0 checkpoint directly from the immutable
revision recorded in its downloader and verifies its SHA-256.
Public validation
An anonymous, token-free download from an empty cache was completed against a
full immutable Hugging Face revision. All four SHA-256 checks passed. The four
downloaded checkpoints then reproduced the expected complete-test accuracies:
58.1% and 59.5% for Server SpikingVGG-11 baseline/TwiST, and 75.9% and 75.7%
for Orin Spikformer baseline/TwiST. See ANONYMOUS_VALIDATION.md.
