{
    "action": "final",
    "content": "## Distribution Summary for Numeric Columns\n\nBased on the statistical analysis of all numeric columns in the dataset, here are the key distributions:\n\n### **Citations**\n- **Range:** 0 to 228,517 citations\n- **Central tendency:** Mean of 4,570 citations, but median of only 493 citations\n- **Spread:** Standard deviation of 15,769 citations\n- **Interpretation:** Highly right-skewed distribution. Most models have relatively few citations (median ~500), but a small number of highly influential models drive the mean much higher. The large standard deviation indicates extreme variability.\n- **Coverage:** 1,485 of 3,574 models (42%) have citation data\n\n### **Parameters**\n- **Range:** 10 to 173.9 trillion parameters\n- **Central tendency:** Mean of 182.5 billion parameters, median of 2.7 billion\n- **Spread:** Standard deviation of 4.2 trillion parameters\n- **Interpretation:** Extremely right-skewed. The median shows most models are in the few-billion parameter range, while a few massive models (like the 173.9T parameter outlier) pull the mean far higher. The enormous standard deviation reflects the presence of both tiny experimental models and frontier-scale systems.\n- **Coverage:** 2,323 of 3,574 models (65%) have parameter counts\n\n### **Training Compute (FLOP)**\n- **Range:** 40 FLOP to 5×10²⁶ FLOP\n- **Central tendency:** Mean of 1.7×10²⁴ FLOP, median of 4.0×10²¹ FLOP\n- **Spread:** Standard deviation of 1.95×10²⁵ FLOP\n- **Interpretation:** Spans ~24 orders of magnitude. The median (4×10²¹ FLOP) represents moderate-scale training, while the mean is pulled up by frontier models requiring 10²⁵–10²⁶ FLOP. This reflects the exponential growth in compute for state-of-the-art models.\n- **Coverage:** 1,402 of 3,574 models (39%) have training compute estimates\n\n### **Training Time (hours)**\n- **Range:** 0.1 to 9,023 hours (~376 days)\n- **Central tendency:** Mean of 471 hours (~20 days), median of 146 hours (~6 days)\n- **Spread:** Standard deviation of 888 hours\n- **Interpretation:** Most models train in under a week (median), but some frontier models require months of continuous training. The distribution is right-skewed with the mean nearly 3× the median.\n- **Coverage:** 554 of 3,574 models (16%) have training time data\n\n### **Epochs**\n- **Range:** 0 to 191,400 epochs\n- **Central tendency:** Mean of 562 epochs, median of 16 epochs\n- **Spread:** Standard deviation of 7,196 epochs\n- **Interpretation:** Extremely right-skewed with massive outliers. The median of 16 epochs is typical for modern large-scale training, while the extreme maximum (191K epochs) likely represents small-scale experimental models trained to convergence on tiny datasets.\n- **Coverage:** 807 of 3,574 models (23%) have epoch counts\n\n### **Finetune Compute (FLOP)**\n- **Range:** 0 to 2.78×10²⁴ FLOP\n- **Central tendency:** Mean of 8.1×10²² FLOP, median of 9.1×10²⁰ FLOP\n- **Spread:** Standard deviation of 3.6×10²³ FLOP\n- **Interpretation:** The median shows typical finetuning uses ~1000× less compute than training from scratch (comparing to training compute median). The mean is ~90× higher than the median, indicating some models undergo extensive finetuning.\n- **Coverage:** 260 of 3,574 models (7%) have finetuning compute data\n\n### **Key Patterns Across All Metrics**\n\n1. **Extreme right skew:** Every numeric column shows mean >> median, indicating distributions dominated by a long tail of high-value outliers\n2. **Sparse coverage:** Citation data exists for only 42% of models, training time for just 16%, and finetuning compute for only 7%\n3. **Scale variation:** Metrics span many orders of magnitude (e.g., 24 orders for training compute), reflecting the dataset's coverage from experimental prototypes to frontier systems\n4. **Standard deviations often exceed means:** This extreme variability is characteristic of exponential growth in AI capabilities and compute investment\n\nThe distributions reveal a field where most models are relatively modest in scale, but a small fraction of frontier models operate at dramatically larger scales across all dimensions."
}