vtava/Laya-v54-optimized-liquid-cenn

Laya V5.4 — Efficient Liquid CeNN Compressed ModernBERT

An experimental 1D token-lattice liquid CeNN, replacing one FFN output projection at original encoder layer 20 with 64 state channels, 2 recurrent steps and depthwise nearest-token-neighbor coupling. All other 28-layer pretrained ModernBERT computation is preserved, with 20 original selectively SVD- compressed attention/FFN projections in the baseline. This is NOT a canonical Chua CeNN or a full Transformer-to-CeNN replacement.

Independent measured TEST (200 typed decisions)

• Original teacher gold accuracy: 76.500%
• Compressed baseline gold accuracy: 75.500%
• Liquid V5.4 gold accuracy: 76.500%
• Liquid change vs compressed baseline: +1.00 percentage points
• Liquid change vs original teacher: +0.00 percentage points
• Teacher decision agreement: 92.000%
• Full FP32 model size: 1,550,800,408 bytes
• Original teacher FP32 size: 1,684,141,004 bytes
• Full model byte saving vs original teacher: 133,340,596
• Projection saving vs compressed baseline: 2,746,620 theoretical FP32 bytes
• Median GPU speed ratio vs compressed baseline: 0.994×
• Minimum GPU speed ratio vs baseline: 0.993×
• All release gates passed: True

This is a compression/fidelity release, not a claim of higher gold accuracy than the teacher. If a speed ratio is below 1×, this model is actually slower on the measured input shapes.

Offline inference

After downloading this model repository once:

```python from modeling_laya_v53 import load_v53_model model,tokenizer=load_v53_model("/local/model",device="cuda") ```

Full FP32 weights, encoder config, tokenizer, standalone loader, and measured statistics are included in this repository. TEST split has been used in previous experiments; independent replication is needed.