Pegi1727/beyond-big-data-: DIINA v1.5.0: Beyond Big Data — Yoruba Case Study & Treasure Corpus`

🌍 Empowering Low-Resource Languages (LRLs) This release introduces the application of DIINA (Dynamic Inhibition-Inspired Neural Architecture) to the Yoruba language, demonstrating a neuro-biologically inspired path for language preservation without the need for "Big Data." ‌ 🧠 Core Innovation: Dynamic Inhibition Unlike standard LLMs that require billions of tokens, DIINA mimics the bilingual inhibitory control mechanism. By dynamically modulating feature activations, the model: Suppresses noise and spurious correlations in small datasets. Focuses on robust, core linguistic structures. Mitigates "catastrophic forgetting" during fine-tuning. ‌ 📊 Empirical Breakthroughs (Case Study: Yoruba) Language ID: 96.8% Accuracy. Perplexity (PPL): Significantly reduced to 98.2, outperforming AfriBERTa and mBERT. Human Evaluation: Native linguists confirmed superior Cultural Appropriateness and Poetic Fluency in generated texts compared to baselines. ‌ 📦 Included in this Release: Architecture Source Code: PyTorch implementation of DynamicInhibitionModule. Yoruba Treasure Corpus (Sample): A curated collection of high-quality Yoruba texts (Folk tales, News, Cultural discourse). Metric Definitions: Comprehensive framework for LI, SLR, IR, and PPL. ‌ Principal Investigator: Dr. Pegah Merrikhi Mission: Moving beyond the "Big Data" paradigm toward equitable language technologies.