# Natural Segmentation Opportunities in the AI Models Dataset Based on the evidence gathered, **yes, there are multiple natural ways to segment this dataset of 3,574 AI models**. The data reveals clear structural dimensions that create meaningful groupings: ## 1. **Domain-Based Segmentation** (Primary Natural Grouping) The **Domain** field provides the most fundamental segmentation: - **Language models**: 1,582 records (44.3%) - the dominant category - **Biology models**: 376 records (10.5%) - **Vision models**: 322 records (9.0%) - **Image generation**: 166 records (4.6%) - **Speech models**: 131 records (3.7%) - **Multimodal combinations**: Multiple categories (e.g., "Multimodal,Language,Vision" with 80 records) This segmentation is natural because it reflects the fundamental application area and data modality each model addresses. ## 2. **Temporal Segmentation** (Evolution-Based) The dataset shows a clear **era-based structure** with dramatically different distributions: - **Pre-2010**: 215 models (6.0%) - **2010-2014**: 132 models (3.7%) - **2015-2019**: 494 models (13.8%) - **2020-2022**: 569 models (15.9%) - **2023+**: 2,147 models (60.1%) - representing the recent AI boom This temporal segmentation is meaningful because it captures distinct phases of AI development, with 60% of all models emerging in the most recent era. ## 3. **Organization Type Segmentation** The **Organization categorization** field reveals institutional patterns: - **Industry-led**: 1,809 models (50.6%) - **Academia-led**: 529 models (14.8%) - **Collaborative efforts**: Various combinations (e.g., "Industry,Academia" with 158 models, 4.4%) - **Research collectives**: 29 models (0.8%) - **Government**: 24 models (0.7%) ## 4. **Scale/Capability Segmentation** **Parameter count tiers** create natural technical groupings: - **Unknown scale**: 1,251 models (35.0%) - **<100M params**: 576 models (16.1%) - **100M-1B params**: 378 models (10.6%) - **1B-10B params**: 624 models (17.5%) - **10B-100B params**: 507 models (14.2%) - **100B+ params**: 238 models (6.7%) The cross-tabulation (step-15) shows this scale dimension **evolves with time**: Pre-2010 models were predominantly <100M parameters (109 of 215), while 2023+ models cluster in the 1B-10B range (499 models) and 10B-100B range (450 models). ## 5. **Accessibility Segmentation** **Model accessibility** creates distinct user-facing segments: - **Unreleased**: 829 models (23.2%) - **Open weights (unrestricted)**: 820 models (22.9%) - **API access only**: 403 models (11.3%) - **Open weights (restricted use)**: 302 models (8.4%) - **Open weights (non-commercial)**: 228 models (6.4%) - **Unknown/blank**: 869 models (24.3%) ## 6. **Multi-Dimensional Segmentation** The evidence shows these dimensions **interact meaningfully**: - **Domain × Approach** (step-12): Language models split between unlabeled (1,452) and self-supervised learning (96), while Vision models show 49 supervised models - **Era × Parameter tier** (step-15): Clear progression from small models in early eras to large models in recent years ## What Remains Uncertain - **Task-level granularity**: While 317 models are tagged "Language modeling" (8.9%), 119 models (3.3%) have blank task fields, limiting task-based segmentation completeness - **Training compute patterns**: 59.7% of models (2,132 records) lack training compute estimation methods, making compute-based segmentation incomplete - **Cross-dimensional relationships**: The investigation reached step limits before fully exploring how accessibility patterns vary by domain, or how organization type correlates with model scale ## Recommendation The **most robust segmentation approach** would use **Domain** as the primary dimension (cleanest distribution, 100% populated per step-2), with **Era** and **Parameter tier** as secondary dimensions to capture technological evolution and capability levels. This three-dimensional framework is fully supported by the data and creates actionable segments for analysis.