Hierarchical Memory SNN v3: Word-Level Japanese Language Model with Nucleus Sampling for Diversity Enhancement

I present v3 of my Hierarchical Memory SNN architecture, focusing on word-level Japanese generation and solving the diversity problem: **v3 Updates:**- Word-level tokenization using MeCab morphological analyzer- Nucleus Sampling breakthrough: Diversity improved from 10% to 100%- Quality-Diversity optimization: T=0.8, top-p=0.9 achieves 83.5% quality with 100% diversity- 10,000 unique sentence corpus for improved generalization- Hybrid spike+membrane readout confirmed as optimal **Key Results (v1 + v2 + v3):**- 1,000,000 neurons with only 4MB RAM + 61MB SSD (v1)- 1000x RAM reduction vs. conventional approaches (v1)- 11D Hypercube: 65% PPL improvement over 10D (v2)- Japanese LLM: 28.7M params, PPL 81.92 (v2)- Word-level generation with 359 vocabulary (v3)- Diversity: 1/10 → 10/10 with Nucleus Sampling (v3)- Quality score: 83.5% with optimal parameters (v3) This enables deployment of diverse, efficient language models on resource-constrained edge devices. Source code: https://github.com/hafufu-stack/snn-llm