SymbolicLight V1: Spike-Gated Dual-Path Language Modeling with High Activation Sparsity and Sub-Billion-Scale Pre-Training Evidence
Paper • 2605.21333 • Published
SymbolicLight V1
中文 | English
This Hugging Face repository hosts the separately distributed model artifacts for SymbolicLight V1, a spike-gated dual-path language-model architecture described in arXiv:2605.21333.
The public source-code and reproducibility repository is:
https://github.com/SymbolicLight-AGI/SymbolicLight-V1
Repository Layout
| Path | Contents |
|---|---|
194M/ |
194M controlled-evidence checkpoints and the 194M tokenizer |
0.8B/ |
0.8B scale-up checkpoint and the 0.8B tokenizer |
LICENSE |
Apache License 2.0 |
NOTICE |
Release boundary and notice |
WEIGHTS_LICENSE.md |
Weight and data-license boundary |
The 194M and 0.8B artifacts are intentionally kept separate. Their tokenizers and checkpoint surfaces are not interchangeable.
194M Artifacts
The 194M/ package contains:
- the SL-BPE 48K tokenizer in
194M/tokenizer/ - SNN controlled-evidence checkpoints
- ablation checkpoints
- GPT-2 baseline checkpoints used for historical comparison
See 194M/README.md for the checkpoint manifest.
0.8B Artifact
The 0.8B/ package contains:
- tokenizer assets in
0.8B/tokenizer/ - cleaned weights-only checkpoint at
0.8B/weights/pytorch/latest.pt
The 0.8B checkpoint is a base pre-training and scale-up artifact. It has not been post-trained into an instruction-following assistant model.
See 0.8B/README.md for file details.
Release Boundary
This repository does not include raw training text, raw validation text, source-level data manifests, source identifiers, download URLs, optimizer state, scaler state, or data-loader state.
Users are responsible for using corpora and downstream data that they have the right to process and redistribute.
License
Unless a file states otherwise, the released tokenizers, documentation, release metadata, and SymbolicLight cleaned weights-only artifacts are released under the Apache License 2.0. Baseline checkpoints included for historical comparison should be treated according to their own source and file-level terms if any. See WEIGHTS_LICENSE.md for the data and checkpoint boundary.
Citation
If you use these artifacts in research, cite the SymbolicLight V1 paper and the public repository:
SymbolicLight V1, arXiv:2605.21333
https://arxiv.org/abs/2605.21333
https://github.com/SymbolicLight-AGI/SymbolicLight-V1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support