```json { "action": "final", "content": "## Dataset Columns: All AI Models\n\nThis dataset contains **61 columns** across 3,574 AI model records. Here's what each column represents:\n\n### **Identifiers & Core Attributes**\n- **Model** (VARCHAR) — The name/identifier of the AI model\n- **Organization** (VARCHAR) — The organization that developed the model\n- **Authors** (VARCHAR) — The researchers/developers who created the model\n- **Hugging Face developer id** (VARCHAR) — Developer identifier on Hugging Face platform\n- **Archived links** (VARCHAR) — Archived reference links for the model\n\n### **Classification & Categorization**\n- **Domain** (VARCHAR) — The application domain (e.g., Vision, Video, Robotics)\n- **Task** (VARCHAR) — Specific task the model performs (e.g., image classification, OCR)\n- **Approach** (VARCHAR) — Training approach (e.g., Supervised, Unsupervised)\n- **Notability criteria** (VARCHAR) — Why this model is notable (e.g., Training cost, Significant use)\n- **Model accessibility** (VARCHAR) — How accessible the model is (e.g., Open, Closed, Unreleased)\n- **Confidence** (VARCHAR) — Confidence level in the data (e.g., High, Medium, Low, Unknown)\n- **Foundation model** (categorical) — Whether this is a foundation model\n- **Frontier model** (categorical) — Whether this is a frontier/cutting-edge model\n- **Open model weights?** (categorical) — Whether model weights are openly available\n- **Organization categorization** (VARCHAR) — Type/category of the developing organization\n- **Possibly over 1e23 FLOP** (categorical) — Flag for extremely large training compute\n\n### **Temporal Information**\n- **Publication date** (DATE) — When the model was published (range: 1950-07-02 to 2026-07-31)\n- **Last modified** (temporal) — When the record was last updated\n- **Training time (hours)** (DOUBLE) — Duration of training in hours (range: 0.1 to 9,022.8 hours)\n\n### **Geographic Information**\n- **Country (of organization)** (VARCHAR) — Country/countries where the organization is based\n\n### **Model Architecture & Scale**\n- **Parameters** (DOUBLE) — Number of model parameters (range: 10 to 173.9 trillion)\n- **Base model** (VARCHAR) — The base/parent model this was built upon\n- **Numerical format** (VARCHAR) — Data type precision used (e.g., FP16, INT8)\n\n### **Training Compute Metrics**\n- **Training compute (FLOP)** (DOUBLE) — Total floating-point operations for training (range: 40 to 5×10²⁶ FLOP)\n- **Training compute lower bound** (measure) — Lower estimate of training compute\n- **Training compute upper bound** (measure) — Upper estimate of training compute\n- **Training compute cost (2023 USD)** (DOUBLE) — Estimated cost in 2023 dollars\n- **Training compute estimation method** (VARCHAR) — How the compute was estimated\n- **Finetune compute (FLOP)** (DOUBLE) — Compute used for fine-tuning (range: 0 to 2.78×10²⁴ FLOP)\n- **Post-training compute (FLOP)** (categorical) — Compute used after initial training\n\n### **Training Infrastructure**\n- **Training hardware** (VARCHAR) — Hardware used for training (e.g., Nvidia H100 80G)\n- **Hardware quantity** (measure) — Number of hardware units used\n- **Hardware utilization (HFU)** (categorical) — Hardware FLOPS utilization rate\n- **Hardware utilization (MFU)** (categorical) — Model FLOPS utilization rate\n- **Training chip-hours** (measure) — Total chip-hours consumed\n- **Training cloud compute vendor** (VARCHAR) — Cloud provider used for training\n- **Training power draw (W)** (measure) — Power consumption during training\n- **Training data center** (categorical) — Data center location\n- **Training code accessibility** (VARCHAR) — Whether training code is available\n\n### **Training Data**\n- **Training dataset size (total)** (VARCHAR) — Size of the training dataset\n- **Batch size** (measure) — Training batch size\n- **Epochs** (DOUBLE) — Number of training epochs (range: 0 to 191,400)\n- **WikiText and Penn Treebank data** (VARCHAR) — Specific benchmark dataset information\n\n### **Impact & Documentation**\n- **Citations** (DOUBLE) — Number of academic citations (range: 0 to 228,517)\n- **Reference** (VARCHAR) — Primary reference/citation for the model\n- **Link** (VARCHAR) — URL to model documentation or repository\n- **Abstract** (VARCHAR) — Description/abstract of the model\n\n### **Notes & Additional Context**\n- **Parameters notes** (VARCHAR) — Additional details about parameter counts\n- **Training compute notes** (VARCHAR) — Detailed notes on compute estimation\n- **Dataset size notes** (VARCHAR) — Additional dataset information\n- **Training time notes** (VARCHAR) — Context about training duration\n- **Finetune compute notes** (VARCHAR) — Details about fine-tuning process\n- **Notability criteria notes** (VARCHAR) — Explanation of why the model is notable\n- **Batch size notes** (VARCHAR) — Additional batch size context\n- **Accessibility notes** (VARCHAR) — Details about model access\n- **Post-training compute notes** (categorical) — Notes on post-training work\n\n### **Data Completeness**\nMany columns have significant missing data:\n- **Citations**: 2,089 nulls (58% missing)\n- **Parameters**: 1,251 nulls (35% missing)\n- **Training compute (FLOP)**: 2,172 nulls (61% missing)\n- **Training time (hours)**: 3,020 nulls (84% missing)\n- **Finetune compute (FLOP)**: 3,314 nulls (93% missing)\n- **Epochs**: 2,767 nulls (77% missing)\n\nThis reflects the reality that comprehensive training details are not always publicly disclosed for AI models." } ```