Remove false #1 SOTA benchmark claims, update actual SOTA numbers

SHD SOTA is 96.41% (SpikCommander), not 91.0%. SSC SOTA is 85.98%,
not 76.4%. Removed all "#1 WORLD", "#1 SOTA", "beats Loihi" claims
from README, HuggingFace model cards, SSC docs, training logs, and
results.json. Updated SOTA references to actual current values.

Author: Henry Barnes

README.md

@@ -12,19 +12,19 @@ All results are reproducible. Clone, install, train, deploy.
 
 | Benchmark | Classes | Architecture | Neuron | Float Acc | Params | vs SOTA |
 |-----------|---------|-------------|--------|-----------|--------|---------|
-| **SHD** | 20 | 700→1536→20 (rec) | adLIF | **91.0%** | 3.47M | **#1** (beat SpikFormer 90.4%) |
-| **SSC** | 35 | 700→1024→512→35 (rec) | adLIF | **76.4%** | 2.31M | **#1** (beat Bittar 74.2%) |
+| **SHD** | 20 | 700→1536→20 (rec) | adLIF | **91.0%** | 3.47M | Competitive |
+| **SSC** | 35 | 700→1024→512→35 (rec) | adLIF | **76.4%** | 2.31M | Competitive |
 | **N-MNIST** | 10 | Conv2D+LIF→10 | LIF | **99.2%** | 691K | -0.47% vs PLIF 99.67% |
 | **GSC-35** | 35 | Delay+rec adLIF+BN | adLIF | *training* | ~1.5M | target: 95%+ |
 
-N3 achieves **#1 on SHD and SSC**, the two most challenging spiking audio benchmarks. All models use adaptive LIF neurons with learnable per-neuron thresholds, surrogate gradient BPTT, and cosine LR scheduling.
+All models use adaptive LIF neurons with learnable per-neuron thresholds, surrogate gradient BPTT, and cosine LR scheduling. Actively improving all benchmarks with learnable delays and advanced training techniques.
 
 ## Results — Catalyst N2
 
 | Benchmark | Classes | Architecture | Neuron | Float Acc | Params | Beats |
 |-----------|---------|-------------|--------|-----------|--------|-------|
-| **SHD** | 20 | 700→1024→20 (rec) | adLIF | **90.7%** | 1.79M | Loihi 1 (89.0%) |
-| **SSC** | 35 | 700→1024→512→35 (rec) | adLIF | **72.1%** | 2.31M | Loihi 2 (69.8%) |
+| **SHD** | 20 | 700→1024→20 (rec) | adLIF | **90.7%** | 1.79M | Competitive |
+| **SSC** | 35 | 700→1024→512→35 (rec) | adLIF | **72.1%** | 2.31M | Competitive |
 | **N-MNIST** | 10 | Conv2D+LIF→10 | LIF | **99.2%** | 466K | — |
 | **GSC KWS** | 12 | 40→512→12 (rec, S2S) | adLIF | **88.0%** | 291K | — |
 
@@ -50,10 +50,10 @@ pip install -e .
 ### Train a benchmark
 
 ```bash
-# SHD (Spiking Heidelberg Digits) — 91.0% with N3 adLIF (#1 SOTA)
+# SHD (Spiking Heidelberg Digits) — 91.0% with N3 adLIF
 python shd/train.py --neuron adlif --hidden 1536 --epochs 200 --device cuda:0 --amp
 
-# SSC (Spiking Speech Commands) — 76.4% with N3 adLIF (#1 SOTA)
+# SSC (Spiking Speech Commands) — 76.4% with N3 adLIF
 python ssc/train.py --hidden1 1024 --hidden2 768 --recurrent2 --epochs 70 --device cuda:0 --amp
 
 # N-MNIST — 99.2% with Conv front-end

TRAINING_LOG.md

@@ -24,14 +24,14 @@
 - **90.7%** at epoch 119/200 (early stopped at 169)
 - Architecture: 700 → 1024 (recurrent adLIF) → 20
 - Key breakthrough: recurrent dropout + activity regularization pushed from 87% to 90.7%
-- Beats Loihi 1 (89.0%), within 0.2% of Loihi 2 (90.9%)
+- Competitive with Loihi 2 (90.9%). Current SOTA: 96.41% (SpikCommander)
 - Full details: `shd/README.md`
 
 ### SSC — Spiking Speech Commands (HEADLINE)
 
 - **72.1%** at epoch 33/200 (crashed at epoch 36, numpy OOM in loader)
 - Architecture: 700 → 1024 (recurrent adLIF) → 512 (adLIF) → 35
-- **Beats Loihi 2 by 2.3 percentage points** (72.1% vs 69.8%)
+- Exceeds Loihi 2 (69.8%) but well below SOTA 85.98% (SpikCommander)
 - Train accuracy still climbing at crash — model likely has more headroom
 - Full details: `ssc/README.md`
 

huggingface/leaderboard/README.md

@@ -22,7 +22,7 @@ Community benchmark comparison for spiking neural networks across neuromorphic h
 
 Covers SHD, SSC, N-MNIST, GSC KWS, and DVS128 Gesture benchmarks. Results from Catalyst N2/N3, Intel Loihi 1/2, BrainChip Akida 2, SpiNNaker 2, and software baselines.
 
-**Catalyst N3 results**: SHD 91.0% (#1 SOTA), SSC 76.4% (#1 SOTA), N-MNIST 99.2%.
+**Catalyst N3 results**: SHD 91.0%, SSC 76.4%, N-MNIST 99.2%.
 
 **Submit your results**: Open a discussion or PR on this Space to add your platform's numbers.
 

huggingface/shd_model_card.md

@@ -33,7 +33,7 @@ model-index:
 
 # Catalyst SHD SNN Benchmark
 
-Spiking Neural Network trained on the Spiking Heidelberg Digits (SHD) dataset using surrogate gradient BPTT. **#1 SOTA** on SHD (91.0%).
+Spiking Neural Network trained on the Spiking Heidelberg Digits (SHD) dataset using surrogate gradient BPTT. Achieves 91.0% on SHD with adaptive LIF neurons.
 
 ## Model Description
 
@@ -47,8 +47,8 @@ Spiking Neural Network trained on the Spiking Heidelberg Digits (SHD) dataset us
 
 | Generation | Architecture | Float Accuracy | Params | vs SOTA |
 |------------|-------------|----------------|--------|---------|
-| **N3** | 700→1536→20 (rec, adLIF) | **91.0%** | 3.47M | **#1** (beat SpikFormer 90.4%) |
-| N2 | 700→1024→20 (rec, adLIF) | 90.7% | 1.79M | Beats Loihi 1 (89.0%) |
+| **N3** | 700→1536→20 (rec, adLIF) | **91.0%** | 3.47M | Competitive |
+| N2 | 700→1024→20 (rec, adLIF) | 90.7% | 1.79M | Competitive |
 
 ## Reproduce
 
@@ -57,7 +57,7 @@ git clone https://github.com/catalyst-neuromorphic/catalyst-benchmarks.git
 cd catalyst-benchmarks
 pip install -e .
 
-# N3 (91.0% — #1 SOTA)
+# N3 (91.0%)
 python shd/train.py --neuron adlif --hidden 1536 --epochs 200 --device cuda:0 --amp
 
 # N2 (90.7%)

huggingface/upload.py

@@ -24,7 +24,7 @@
         'task_name': 'Spoken Digit Classification',
         'dataset': 'Spiking Heidelberg Digits (SHD)',
         'architecture': '700 → 1536 (recurrent adLIF) → 20',
-        'description': 'Spiking Neural Network for spoken digit classification on SHD. #1 SOTA (91.0%), beating SpikFormer (90.4%) and Loihi 2 (90.9%).',
+        'description': 'Spiking Neural Network for spoken digit classification on SHD. Achieves 91.0% with adaptive LIF neurons.',
         'checkpoint': '../n3_results/checkpoints/shd_91.0.pt',
         'float_acc': 91.0,
         'quant_acc': None,
@@ -38,7 +38,7 @@
         'task_name': 'Spoken Command Classification',
         'dataset': 'Spiking Speech Commands (SSC)',
         'architecture': '700 → 1024 (recurrent adLIF) → 512 (adLIF) → 35',
-        'description': 'Spiking Neural Network for spoken command classification on SSC. #1 SOTA (76.4%), beating Bittar (74.2%) and Loihi 2 (69.8%).',
+        'description': 'Spiking Neural Network for spoken command classification on SSC. Achieves 76.4% with adaptive LIF neurons.',
         'checkpoint': '../n3_results/checkpoints/ssc_76.4.pt',
         'float_acc': 76.4,
         'quant_acc': None,

n3_results/N3_BENCHMARKS.md

@@ -4,8 +4,8 @@
 
 | Benchmark | Accuracy | SOTA | vs SOTA | Params | Architecture | Status |
 |-----------|----------|------|---------|--------|--------------|--------|
-| SHD | **91.0%** | 90.4% (SpikFormer) | **+0.6%** | 3.47M | 700->1536(rec adLIF)->35 | DONE, #1 |
-| SSC | **76.4%** | 74.2% (Bittar) | **+2.2%** | 2.31M | 700->1024(rec adLIF)->512(adLIF)->35 | DONE, #1 |
+| SHD | **91.0%** | 96.41% (SpikCommander) | -5.4% | 3.47M | 700->1536(rec adLIF)->35 | DONE |
+| SSC | **76.4%** | 85.98% (SpikCommander) | -9.6% | 2.31M | 700->1024(rec adLIF)->512(adLIF)->35 | DONE |
 | N-MNIST | **99.2%** | 99.67% (PLIF) | -0.47% | 691K | Conv SNN (LIF) | DONE |
 | GSC-35 | 43.9% | 95.69% (d-cAdLIF) | -51.8% | ~1.5M | Delay+rec adLIF+BN (v4) | IN PROGRESS |
 | N-Caltech101 | 41.1% | 84.35% (EV-VGCNN) | -43.3% | 1.47M | FC rec adLIF | PAUSED (resumable) |

results.json

@@ -350,6 +350,39 @@
     "hidden": "1024+512",
     "date": "2026-03-05",
     "training_time_s": 40000,
-    "notes": "SSC N3 v3 final. 700->1024(rec adLIF)->512(adLIF)->35. #1 SOTA (beat Bittar 74.2% by 2.2%). Warm-started from epoch 33 checkpoint, converged at 76.4%."
+    "notes": "SSC N3 v3 final. 700->1024(rec adLIF)->512(adLIF)->35. Warm-started from epoch 33 checkpoint, converged at 76.4%. Current SOTA: 85.98% (SpikCommander)."
+  },
+  {
+    "benchmark": "kan-bspline",
+    "accuracy_float": 79.37,
+    "n_params": 1708052,
+    "epochs": 100,
+    "best_epoch": 52,
+    "neuron_type": "lif",
+    "hidden": null,
+    "date": "2026-03-06",
+    "training_time_s": 10713.2
+  },
+  {
+    "benchmark": "shd-attention",
+    "accuracy_float": 86.79,
+    "n_params": 1682452,
+    "epochs": 100,
+    "best_epoch": 31,
+    "neuron_type": "adlif",
+    "hidden": 512,
+    "date": "2026-03-06",
+    "training_time_s": 11717.3
+  },
+  {
+    "benchmark": "dvs_gesture",
+    "accuracy_float": 80.3,
+    "n_params": 466529,
+    "epochs": 200,
+    "best_epoch": 89,
+    "neuron_type": "conv+adlif",
+    "hidden": null,
+    "date": "2026-03-06",
+    "training_time_s": 16057.8
   }
 ]

ssc/README.md

@@ -1,6 +1,6 @@
 # SSC — Spiking Speech Commands
 
-**Result: 72.1% test accuracy** (beats Intel Loihi 2 at 69.8%)
+**Result: 72.1% test accuracy**
 
 ## Benchmark Details
 
@@ -94,4 +94,4 @@ Train accuracy was still climbing at crash time, suggesting the model could reac
 | Intel Loihi 2 | 69.8% | Bittar & Bhatt 2024 |
 | Software (Bittar) | 74.2% | Bittar & Bhatt 2024 |
 
-**Catalyst N2 beats Loihi 2 by 2.3 percentage points on SSC.**
+Current SOTA is 85.98% (SpikCommander, 2025). We are actively improving this result with learnable delays.

ssc/train.py

@@ -5,7 +5,7 @@
   - NEW: Recurrence on BOTH layers, recurrent dropout on layer 1
   - NEW: Activity regularization
   - Architecture: 700 -> 1024 (rec adLIF) -> 768 (rec adLIF) -> 35 (readout)
-  - Target: 77%+ float (beats Loihi 2's 69.8% by 7%+)
+  - Target: 85%+ float (current SOTA: 85.98% SpikCommander)
 
 Usage:
     python ssc/train.py --data-dir data/ssc --epochs 300 --amp --device cuda:0