Upload README.md with huggingface_hub

README.md

@@ -1,122 +1,93 @@
 ---
-license: apache-2.0
-language:
-  - de
-  - en
+language: en
+license: mit
+library_name: pytorch
 tags:
   - spiking-neural-network
   - neuromorphic
-  - audio-classification
-  - SHD
-  - LIF
-  - FPGA
+  - surrogate-gradient
+  - benchmark
+  - catalyst
+  - shd
 datasets:
-  - custom
+  - shd
 metrics:
   - accuracy
 model-index:
-  - name: Catalyst SHD SNN
+  - name: Catalyst SHD SNN Benchmark
     results:
       - task:
           type: audio-classification
           name: Spoken Digit Classification
         dataset:
           name: Spiking Heidelberg Digits (SHD)
-          type: custom
+          type: shd
         metrics:
-          - name: Accuracy (float32)
+          - name: Float Accuracy
             type: accuracy
-            value: 85.9
-          - name: Accuracy (int16 quantized)
+            value: 90.68
+          - name: Quantized Accuracy (int16)
             type: accuracy
-            value: 85.4
+            value: 90.2
 ---
 
-# Catalyst SHD Spiking Neural Network
+# Catalyst SHD SNN Benchmark
 
-## Model Description
-
-A recurrent spiking neural network (SNN) trained on the Spiking Heidelberg Digits (SHD) spoken digit classification benchmark, designed for deployment on Catalyst neuromorphic processors.
+Spiking Neural Network for spoken digit classification on SHD.
 
-**Architecture**: 700 → 512 (recurrent, LIF) → 20
+## Model Description
 
-| Parameter | Value |
-|-----------|-------|
-| Input channels | 700 (cochlea model spike trains) |
-| Hidden neurons | 512 (recurrent LIF) |
-| Output classes | 20 (German + English digits 0-9) |
-| Trainable parameters | ~421,000 |
-| Neuron model | Leaky Integrate-and-Fire (CUBA) |
-| Surrogate gradient | Fast sigmoid (scale=25) |
-| Time bins | 250 (4ms bins, 1s window) |
+- **Architecture**: 700 → 1024 (recurrent adLIF) → 20
+- **Neuron model**: Adaptive Leaky Integrate-and-Fire (adLIF) with Symplectic Euler discretization
+- **Training**: Surrogate gradient BPTT, fast-sigmoid surrogate (scale=25)
+- **Hardware target**: Catalyst N1/N2/N3 neuromorphic processors
+- **Quantization**: Float32 weights -> int16, membrane decay -> 12-bit fixed-point
 
-## Performance
+## Results
 
 | Metric | Value |
 |--------|-------|
-| Float accuracy (best) | **85.9%** |
-| Quantized accuracy (16-bit) | **85.4%** |
-| Quantization loss | 0.4% |
-
-Surpasses Cramer et al. (2020) at 83.2% and Zenke & Vogels (2021) at 83.4%.
-
-## Training
-
-- **Optimizer**: AdamW (lr=1e-3, weight_decay=1e-4)
-- **Scheduler**: CosineAnnealingLR
-- **Epochs**: 200
-- **Batch size**: 128
-- **Gradient clipping**: norm=1.0
-- **Dropout**: 0.3
+| Float accuracy | 90.68% |
+| Quantized accuracy (int16) | 90.2% |
+| Parameters | 1,789,972 |
+| Quantization loss | 0.5% |
 
-## Hardware Deployment
+## Reproduce
 
-Trained for deployment on Catalyst N1/N2 neuromorphic processors:
-- Weight quantization: float → int16 (scale = threshold_hw / threshold_float)
-- Membrane decay: beta → decay_v = round(beta * 4096) [12-bit fractional]
-- Target hardware threshold: 1000
-- Deployment via: `python shd_deploy.py --checkpoint shd_model.pt`
-
-## Files
-
-- `shd_model.pt` — PyTorch checkpoint (float32 weights)
-- `config.json` — Model architecture and training config
+```bash
+git clone https://github.com/catalyst-neuromorphic/catalyst-benchmarks.git
+cd catalyst-benchmarks
+pip install -e .
+python shd/train.py --device cuda:0
+```
 
-## Usage
+## Deploy to Catalyst Hardware
 
 ```python
-# Install SDK
-pip install catalyst-cloud
-
-# Or use directly with PyTorch
-import torch
-checkpoint = torch.load("shd_model.pt")
+import catalyst_cloud
+
+client = catalyst_cloud.Client()
+result = client.simulate(
+    model="catalyst-neuromorphic/shd-snn-benchmark",
+    input_data=your_spikes,
+    processor="n2"
+)
 ```
 
-For FPGA deployment, use the Catalyst SDK:
-```bash
-pip install catalyst-neurocore
-python -c "from neurocore import Network; print('SDK ready')"
-```
-
-## Dataset
+## Links
 
-[Spiking Heidelberg Digits (SHD)](https://zenkelab.org/resources/spiking-heidelberg-digits-shd/)
-- ~8,000 training samples, ~2,600 test samples
-- 700-channel cochlea model spike recordings of spoken digits
-- 20 classes (digits 0-9 in German and English)
+- **Benchmark repo**: [catalyst-neuromorphic/catalyst-benchmarks](https://github.com/catalyst-neuromorphic/catalyst-benchmarks)
+- **Cloud API**: [catalyst-neuromorphic.com](https://catalyst-neuromorphic.com)
+- **N2 paper**: [Zenodo DOI 10.5281/zenodo.18728256](https://zenodo.org/records/18728256)
+- **N1 paper**: [Zenodo DOI 10.5281/zenodo.18727094](https://zenodo.org/records/18727094)
 
 ## Citation
 
-[https://doi.org/10.5281/zenodo.18727094]
-
-## License
-
-Model weights: Apache 2.0
-SDK: BSL 1.1 (source-available, free for research)
-
-## Links
-
-- [GitHub](https://github.com/catalyst-neuromorphic/catalyst-neurocore)
-- [Cloud API](https://catalyst-neuromorphic.com/cloud)
-- [Website](https://catalyst-neuromorphic.com)
+```bibtex
+@misc{catalyst-benchmarks-2026,
+  author = {Shulayev Barnes, Henry},
+  title = {Catalyst Neuromorphic Benchmarks},
+  year = {2026},
+  url = {https://github.com/catalyst-neuromorphic/catalyst-benchmarks}
+}
+```