{
  "name": "epoch-data-on-ai-models",
  "title": "Epoch Data on AI Models",
  "description": "Comprehensive database of over 2800 AI/ML models tracking key factors driving machine learning progress, including parameters, training compute, training dataset size, publication date, organization, and more. Sourced from Epoch AI.",
  "licenses": [
    {
      "name": "CC-BY-4.0",
      "title": "Creative Commons Attribution 4.0",
      "path": "https://creativecommons.org/licenses/by/4.0/"
    }
  ],
  "sources": [
    {
      "title": "Epoch AI — Notable AI Models",
      "path": "https://epochai.org/data/notable-ai-models"
    }
  ],
  "views": [
    {
      "name": "model-scale-over-time",
      "title": "Training Compute of Notable AI Models Over Time",
      "description": "Each dot is a notable AI model plotted by its training compute (FLOP). The typical model used ~1.5× more compute per year before 2010, accelerating to ~3.8× per year through the Deep Learning Era (2010–2022). Since 2023, the pace has jumped dramatically — the median model's training compute grew ~44× from 2023 to 2024 alone, far above the prior baseline.",
      "resources": ["notable-ai-models"],
      "specType": "plot",
      "spec": {
        "dateFields": ["Publication date"],
        "height": 480,
        "marginLeft": 90,
        "x": { "label": null },
        "y": { "label": "Training compute (FLOP)", "type": "log", "domain": [1e8, 1e27] },
        "color": { "legend": false },
        "computed": [
          { "field": "PrimaryDomain", "from": "Domain", "extract": "firstToken" }
        ],
        "filter": { "field": "Training compute (FLOP)", "notNull": true },
        "marks": [
          {
            "type": "ruleX",
            "staticData": [{ "x": "2010-01-01" }],
            "x": "x",
            "stroke": "#9ca3af",
            "strokeDasharray": "4,4",
            "tip": false
          },
          {
            "type": "ruleX",
            "staticData": [{ "x": "2023-01-01" }],
            "x": "x",
            "stroke": "#9ca3af",
            "strokeDasharray": "4,4",
            "tip": false
          },
          {
            "type": "dot",
            "x": "Publication date",
            "y": "Training compute (FLOP)",
            "stroke": "PrimaryDomain",
            "r": 3,
            "opacity": 0.45,
            "tip": true
          },
          {
            "type": "line",
            "staticData": [
              { "x": "1952-01-01", "y": 5e6 },
              { "x": "2010-01-01", "y": 1e17 }
            ],
            "x": "x", "y": "y",
            "stroke": "#9ca3af",
            "strokeDasharray": "6,4",
            "tip": false
          },
          {
            "type": "line",
            "staticData": [
              { "x": "2010-01-01", "y": 1e17 },
              { "x": "2023-01-01", "y": 1.52e22 }
            ],
            "x": "x", "y": "y",
            "stroke": "#374151",
            "strokeDasharray": "6,4",
            "tip": false
          },
          {
            "type": "line",
            "staticData": [
              { "x": "2023-01-01", "y": 1.52e22 },
              { "x": "2025-06-01", "y": 3.86e25 }
            ],
            "x": "x", "y": "y",
            "stroke": "#CC503E",
            "strokeDasharray": "6,4",
            "tip": false
          },
          {
            "type": "text",
            "staticData": [{ "x": "2011-01-01", "y": 3e8, "label": "Deep Learning Era (~3.8× / year)" }],
            "x": "x", "y": "y", "text": "label",
            "stroke": "#374151",
            "fontSize": 11,
            "textAnchor": "start"
          }
        ]
      }
    }
  ],
  "_views_vega_lite_archived": [
    {
      "name": "parameters-over-time",
      "title": "Model Parameters Over Time (2015–2025)",
      "resources": ["notable-ai-models"],
      "specType": "vega-lite",
      "spec": {
        "$schema": "https://vega.github.io/schema/vega-lite/v5.json",
        "transform": [
          {
            "filter": "year(datum['Publication date']) >= 2015 && toNumber(datum['Parameters']) > 0"
          }
        ],
        "layer": [
          {
            "mark": { "type": "point", "filled": true, "size": 60, "opacity": 0.7 },
            "encoding": {
              "x": {
                "field": "Publication date",
                "type": "temporal",
                "title": "Publication Date",
                "axis": { "grid": false }
              },
              "color": {
                "field": "Domain",
                "type": "nominal",
                "legend": { "title": "Domain" }
              },
              "tooltip": [
                { "field": "Model", "type": "nominal" },
                { "field": "Organization", "type": "nominal" },
                { "field": "Publication date", "type": "temporal", "title": "Date" },
                { "field": "Parameters", "type": "quantitative", "format": ".3s" },
                { "field": "Domain", "type": "nominal" }
              ],
              "y": {
                "field": "Parameters",
                "type": "quantitative",
                "scale": { "type": "log" },
                "title": "Parameters (log scale)",
                "axis": { "grid": false }
              }
            }
          }
        ]
      }
    }
  ],
  "resources": [
    {
      "path": "data/all_ai_models.csv",
      "name": "all-ai-models",
      "title": "All AI Models",
      "description": "All AI models in the Epoch database (~21,000 entries).",
      "mediatype": "text/csv",
      "schema": {
        "fields": [
          {"name": "Model", "type": "string", "description": "Name of the AI model"},
          {"name": "Domain", "type": "string", "description": "Domain(s) the model operates in (e.g. Language, Vision)"},
          {"name": "Task", "type": "string", "description": "Task(s) the model is designed for"},
          {"name": "Organization", "type": "string", "description": "Organization(s) that developed the model"},
          {"name": "Authors", "type": "string", "description": "Authors of the model or associated paper"},
          {"name": "Publication date", "type": "date", "format": "%Y-%m-%d", "description": "Date the model was published or released"},
          {"name": "Reference", "type": "string", "description": "Citation reference for the model"},
          {"name": "Link", "type": "string", "description": "URL to model paper or announcement"},
          {"name": "Citations", "type": "number", "description": "Number of citations"},
          {"name": "Notability criteria", "type": "string", "description": "Criteria that make this model notable"},
          {"name": "Notability criteria notes", "type": "string", "description": "Additional notes on notability criteria"},
          {"name": "Parameters", "type": "number", "description": "Number of model parameters"},
          {"name": "Parameters notes", "type": "string", "description": "Notes on parameter count"},
          {"name": "Training compute (FLOP)", "type": "number", "description": "Total training compute in floating point operations"},
          {"name": "Training compute notes", "type": "string", "description": "Notes on training compute estimate"},
          {"name": "Training dataset", "type": "string", "description": "Name or description of the training dataset"},
          {"name": "Training dataset notes", "type": "string", "description": "Notes on the training dataset"},
          {"name": "Training dataset size (datapoints)", "type": "number", "description": "Number of datapoints in the training dataset"},
          {"name": "Dataset size notes", "type": "string", "description": "Notes on dataset size"},
          {"name": "Training time (hours)", "type": "number", "description": "Total training time in hours"},
          {"name": "Training time notes", "type": "string", "description": "Notes on training time estimate"},
          {"name": "Training hardware", "type": "string", "description": "Hardware used for training (e.g. A100, H100)"},
          {"name": "Approach", "type": "string", "description": "Modeling approach or architecture type"},
          {"name": "Confidence", "type": "string", "description": "Confidence level of the data entries"},
          {"name": "Abstract", "type": "string", "description": "Abstract of the associated paper"},
          {"name": "Epochs", "type": "number", "description": "Number of training epochs"},
          {"name": "Benchmark data", "type": "string", "description": "Benchmark evaluation data"},
          {"name": "Model accessibility", "type": "string", "description": "Accessibility of the model weights (e.g. open, closed)"},
          {"name": "Country (of organization)", "type": "string", "description": "Country where the developing organization is based"},
          {"name": "Base model", "type": "string", "description": "Base model this model was fine-tuned from, if any"},
          {"name": "Finetune compute (FLOP)", "type": "number", "description": "Compute used for fine-tuning in FLOP"},
          {"name": "Finetune compute notes", "type": "string", "description": "Notes on fine-tune compute estimate"},
          {"name": "Hardware quantity", "type": "number", "description": "Number of hardware units used for training"},
          {"name": "Hardware utilization (MFU)", "type": "number", "description": "Model FLOP utilization (MFU) of training hardware"},
          {"name": "Last modified", "type": "string", "description": "Timestamp when the record was last modified"},
          {"name": "Training cloud compute vendor", "type": "string", "description": "Cloud vendor used for training compute"},
          {"name": "Training data center", "type": "string", "description": "Data center used for training"},
          {"name": "Archived links", "type": "string", "description": "Archived URLs for the model or paper"},
          {"name": "Batch size", "type": "number", "description": "Training batch size"},
          {"name": "Batch size notes", "type": "string", "description": "Notes on batch size"},
          {"name": "Organization categorization", "type": "string", "description": "Category of the developing organization (e.g. Industry, Academia)"},
          {"name": "Foundation model", "type": "boolean", "description": "Whether this is a foundation model"},
          {"name": "Training compute lower bound", "type": "number", "description": "Lower bound estimate of training compute in FLOP"},
          {"name": "Training compute upper bound", "type": "number", "description": "Upper bound estimate of training compute in FLOP"},
          {"name": "Training chip-hours", "type": "number", "description": "Total chip-hours used for training"},
          {"name": "Training code accessibility", "type": "string", "description": "Accessibility of training code"},
          {"name": "Accessibility notes", "type": "string", "description": "Notes on accessibility of model or code"},
          {"name": "Organization categorization (from Organization)", "type": "string", "description": "Organization category derived from organization field"},
          {"name": "Possibly over 1e23 FLOP", "type": "boolean", "description": "Whether training compute may exceed 1e23 FLOP"},
          {"name": "Training compute cost (2023 USD)", "type": "number", "description": "Estimated training compute cost in 2023 US dollars"},
          {"name": "Utilization notes", "type": "string", "description": "Notes on hardware utilization"},
          {"name": "Numerical format", "type": "string", "description": "Numerical precision format used in training (e.g. FP16, BF16)"},
          {"name": "Frontier model", "type": "boolean", "description": "Whether this model was a frontier model at the time of release"},
          {"name": "Training power draw (W)", "type": "number", "description": "Power consumption during training in watts"},
          {"name": "Training compute estimation method", "type": "string", "description": "Method used to estimate training compute"},
          {"name": "Hugging Face developer id", "type": "string", "description": "Hugging Face developer or organization identifier"},
          {"name": "Post-training compute (FLOP)", "type": "number", "description": "Compute used for post-training (RLHF, fine-tuning, etc.) in FLOP"},
          {"name": "Post-training compute notes", "type": "string", "description": "Notes on post-training compute estimate"},
          {"name": "Hardware utilization (HFU)", "type": "number", "description": "Hardware FLOP utilization (HFU) during training"}
        ]
      }
    },
    {
      "path": "data/notable_ai_models.csv",
      "name": "notable-ai-models",
      "title": "Notable AI Models",
      "description": "Subset of notable AI models with richer metadata (~7,400 entries).",
      "mediatype": "text/csv",
      "schema": {
        "fields": [
          {"name": "Model", "type": "string", "description": "Name of the AI model"},
          {"name": "Organization", "type": "string", "description": "Organization(s) that developed the model"},
          {"name": "Publication date", "type": "date", "format": "%Y-%m-%d", "description": "Date the model was published or released"},
          {"name": "Domain", "type": "string", "description": "Domain(s) the model operates in (e.g. Language, Vision)"},
          {"name": "Task", "type": "string", "description": "Task(s) the model is designed for"},
          {"name": "Parameters", "type": "number", "description": "Number of model parameters"},
          {"name": "Parameters notes", "type": "string", "description": "Notes on parameter count"},
          {"name": "Training compute (FLOP)", "type": "number", "description": "Total training compute in floating point operations"},
          {"name": "Training compute notes", "type": "string", "description": "Notes on training compute estimate"},
          {"name": "Training dataset", "type": "string", "description": "Name or description of the training dataset"},
          {"name": "Training dataset size (datapoints)", "type": "number", "description": "Number of datapoints in the training dataset"},
          {"name": "Dataset size notes", "type": "string", "description": "Notes on dataset size"},
          {"name": "Confidence", "type": "string", "description": "Confidence level of the data entries"},
          {"name": "Link", "type": "string", "description": "URL to model paper or announcement"},
          {"name": "Reference", "type": "string", "description": "Citation reference for the model"},
          {"name": "Citations", "type": "number", "description": "Number of citations"},
          {"name": "Authors", "type": "string", "description": "Authors of the model or associated paper"},
          {"name": "Abstract", "type": "string", "description": "Abstract of the associated paper"},
          {"name": "Organization categorization", "type": "string", "description": "Category of the developing organization (e.g. Industry, Academia)"},
          {"name": "Country (of organization)", "type": "string", "description": "Country where the developing organization is based"},
          {"name": "Notability criteria", "type": "string", "description": "Criteria that make this model notable"},
          {"name": "Notability criteria notes", "type": "string", "description": "Additional notes on notability criteria"},
          {"name": "Epochs", "type": "number", "description": "Number of training epochs"},
          {"name": "Training time (hours)", "type": "number", "description": "Total training time in hours"},
          {"name": "Training time notes", "type": "string", "description": "Notes on training time estimate"},
          {"name": "Training hardware", "type": "string", "description": "Hardware used for training (e.g. A100, H100)"},
          {"name": "Hardware quantity", "type": "number", "description": "Number of hardware units used for training"},
          {"name": "Hardware utilization (MFU)", "type": "number", "description": "Model FLOP utilization (MFU) of training hardware"},
          {"name": "Training compute cost (2023 USD)", "type": "number", "description": "Estimated training compute cost in 2023 US dollars"},
          {"name": "Compute cost notes", "type": "string", "description": "Notes on compute cost estimate"},
          {"name": "Training power draw (W)", "type": "number", "description": "Power consumption during training in watts"},
          {"name": "Base model", "type": "string", "description": "Base model this model was fine-tuned from, if any"},
          {"name": "Finetune compute (FLOP)", "type": "number", "description": "Compute used for fine-tuning in FLOP"},
          {"name": "Finetune compute notes", "type": "string", "description": "Notes on fine-tune compute estimate"},
          {"name": "Batch size", "type": "number", "description": "Training batch size"},
          {"name": "Batch size notes", "type": "string", "description": "Notes on batch size"},
          {"name": "Model accessibility", "type": "string", "description": "Accessibility of the model weights (e.g. open, closed)"},
          {"name": "Training code accessibility", "type": "string", "description": "Accessibility of training code"},
          {"name": "Inference code accessibility", "type": "string", "description": "Accessibility of inference code"},
          {"name": "Accessibility notes", "type": "string", "description": "Notes on accessibility of model or code"},
          {"name": "Numerical format", "type": "string", "description": "Numerical precision format used in training (e.g. FP16, BF16)"},
          {"name": "Frontier model", "type": "boolean", "description": "Whether this model was a frontier model at the time of release"},
          {"name": "Hardware acquisition cost", "type": "number", "description": "Cost of acquiring the training hardware in USD"},
          {"name": "Hardware utilization (HFU)", "type": "number", "description": "Hardware FLOP utilization (HFU) during training"},
          {"name": "Training compute cost (cloud)", "type": "number", "description": "Estimated training compute cost using cloud pricing in USD"},
          {"name": "Training compute cost (upfront)", "type": "number", "description": "Estimated training compute cost using upfront hardware pricing in USD"}
        ]
      }
    },
    {
      "path": "data/large_scale_ai_models.csv",
      "name": "large-scale-ai-models",
      "title": "Large-Scale AI Models",
      "description": "Large-scale AI models subset (~3,600 entries).",
      "mediatype": "text/csv",
      "schema": {
        "fields": [
          {"name": "Model", "type": "string", "description": "Name of the AI model"},
          {"name": "Domain", "type": "string", "description": "Domain(s) the model operates in (e.g. Language, Vision)"},
          {"name": "Task", "type": "string", "description": "Task(s) the model is designed for"},
          {"name": "Authors", "type": "string", "description": "Authors of the model or associated paper"},
          {"name": "Model accessibility", "type": "string", "description": "Accessibility of the model weights (e.g. open, closed)"},
          {"name": "Link", "type": "string", "description": "URL to model paper or announcement"},
          {"name": "Citations", "type": "number", "description": "Number of citations"},
          {"name": "Reference", "type": "string", "description": "Citation reference for the model"},
          {"name": "Publication date", "type": "date", "format": "%Y-%m-%d", "description": "Date the model was published or released"},
          {"name": "Organization", "type": "string", "description": "Organization(s) that developed the model"},
          {"name": "Parameters", "type": "number", "description": "Number of model parameters"},
          {"name": "Parameters notes", "type": "string", "description": "Notes on parameter count"},
          {"name": "Training compute (FLOP)", "type": "number", "description": "Total training compute in floating point operations"},
          {"name": "Training compute notes", "type": "string", "description": "Notes on training compute estimate"},
          {"name": "Training dataset", "type": "string", "description": "Name or description of the training dataset"},
          {"name": "Training dataset notes", "type": "string", "description": "Notes on the training dataset"},
          {"name": "Training dataset size (datapoints)", "type": "number", "description": "Number of datapoints in the training dataset"},
          {"name": "Dataset size notes", "type": "string", "description": "Notes on dataset size"},
          {"name": "Training time (hours)", "type": "number", "description": "Total training time in hours"},
          {"name": "Training time notes", "type": "string", "description": "Notes on training time estimate"},
          {"name": "Training hardware", "type": "string", "description": "Hardware used for training (e.g. A100, H100)"},
          {"name": "Confidence", "type": "string", "description": "Confidence level of the data entries"},
          {"name": "Abstract", "type": "string", "description": "Abstract of the associated paper"},
          {"name": "Country (of organization)", "type": "string", "description": "Country where the developing organization is based"},
          {"name": "Base model", "type": "string", "description": "Base model this model was fine-tuned from, if any"},
          {"name": "Finetune compute (FLOP)", "type": "number", "description": "Compute used for fine-tuning in FLOP"},
          {"name": "Finetune compute notes", "type": "string", "description": "Notes on fine-tune compute estimate"},
          {"name": "Hardware quantity", "type": "number", "description": "Number of hardware units used for training"},
          {"name": "Hardware utilization (MFU)", "type": "number", "description": "Model FLOP utilization (MFU) of training hardware"},
          {"name": "Training code accessibility", "type": "string", "description": "Accessibility of training code"},
          {"name": "Accessibility notes", "type": "string", "description": "Notes on accessibility of model or code"},
          {"name": "Organization categorization (from Organization)", "type": "string", "description": "Organization category derived from organization field"},
          {"name": "Hardware utilization (HFU)", "type": "number", "description": "Hardware FLOP utilization (HFU) during training"},
          {"name": "Training compute cost (cloud)", "type": "number", "description": "Estimated training compute cost using cloud pricing in USD"},
          {"name": "Training compute cost (upfront)", "type": "number", "description": "Estimated training compute cost using upfront hardware pricing in USD"}
        ]
      }
    },
    {
      "path": "data/frontier_ai_models.csv",
      "name": "frontier-ai-models",
      "title": "Frontier AI Models",
      "description": "Frontier AI models subset — the most capable models at each point in time (~1,600 entries).",
      "mediatype": "text/csv",
      "schema": {
        "fields": [
          {"name": "Model", "type": "string", "description": "Name of the AI model"},
          {"name": "Domain", "type": "string", "description": "Domain(s) the model operates in (e.g. Language, Vision)"},
          {"name": "Task", "type": "string", "description": "Task(s) the model is designed for"},
          {"name": "Authors", "type": "string", "description": "Authors of the model or associated paper"},
          {"name": "Notability criteria", "type": "string", "description": "Criteria that make this model notable"},
          {"name": "Notability criteria notes", "type": "string", "description": "Additional notes on notability criteria"},
          {"name": "Model accessibility", "type": "string", "description": "Accessibility of the model weights (e.g. open, closed)"},
          {"name": "Link", "type": "string", "description": "URL to model paper or announcement"},
          {"name": "Citations", "type": "number", "description": "Number of citations"},
          {"name": "Reference", "type": "string", "description": "Citation reference for the model"},
          {"name": "Publication date", "type": "date", "format": "%Y-%m-%d", "description": "Date the model was published or released"},
          {"name": "Organization", "type": "string", "description": "Organization(s) that developed the model"},
          {"name": "Parameters", "type": "number", "description": "Number of model parameters"},
          {"name": "Parameters notes", "type": "string", "description": "Notes on parameter count"},
          {"name": "Training compute (FLOP)", "type": "number", "description": "Total training compute in floating point operations"},
          {"name": "Training compute notes", "type": "string", "description": "Notes on training compute estimate"},
          {"name": "Training dataset", "type": "string", "description": "Name or description of the training dataset"},
          {"name": "Training dataset notes", "type": "string", "description": "Notes on the training dataset"},
          {"name": "Training dataset size (datapoints)", "type": "number", "description": "Number of datapoints in the training dataset"},
          {"name": "Dataset size notes", "type": "string", "description": "Notes on dataset size"},
          {"name": "Epochs", "type": "number", "description": "Number of training epochs"},
          {"name": "Inference compute (FLOP)", "type": "number", "description": "Compute per inference pass in FLOP"},
          {"name": "Inference compute notes", "type": "string", "description": "Notes on inference compute estimate"},
          {"name": "Training time (hours)", "type": "number", "description": "Total training time in hours"},
          {"name": "Training time notes", "type": "string", "description": "Notes on training time estimate"},
          {"name": "Training hardware", "type": "string", "description": "Hardware used for training (e.g. A100, H100)"},
          {"name": "Approach", "type": "string", "description": "Modeling approach or architecture type"},
          {"name": "Compute cost notes", "type": "string", "description": "Notes on compute cost estimate"},
          {"name": "Compute sponsor categorization", "type": "string", "description": "Category of the compute sponsor"},
          {"name": "Confidence", "type": "string", "description": "Confidence level of the data entries"},
          {"name": "Abstract", "type": "string", "description": "Abstract of the associated paper"},
          {"name": "Last modified", "type": "string", "description": "Timestamp when the record was last modified"},
          {"name": "Created By", "type": "string", "description": "Person who created this record"},
          {"name": "Benchmark data", "type": "string", "description": "Benchmark evaluation data"},
          {"name": "Exclude", "type": "boolean", "description": "Whether this model is excluded from certain analyses"},
          {"name": "Country (of organization)", "type": "string", "description": "Country where the developing organization is based"},
          {"name": "Base model", "type": "string", "description": "Base model this model was fine-tuned from, if any"},
          {"name": "Finetune compute (FLOP)", "type": "number", "description": "Compute used for fine-tuning in FLOP"},
          {"name": "Finetune compute notes", "type": "string", "description": "Notes on fine-tune compute estimate"},
          {"name": "Hardware quantity", "type": "number", "description": "Number of hardware units used for training"},
          {"name": "Hardware utilization (MFU)", "type": "number", "description": "Model FLOP utilization (MFU) of training hardware"},
          {"name": "Training cost trends", "type": "string", "description": "Trend information for training costs"},
          {"name": "Training cloud compute vendor", "type": "string", "description": "Cloud vendor used for training compute"},
          {"name": "Training data center", "type": "string", "description": "Data center used for training"},
          {"name": "Archived links", "type": "string", "description": "Archived URLs for the model or paper"},
          {"name": "Batch size", "type": "number", "description": "Training batch size"},
          {"name": "Batch size notes", "type": "string", "description": "Notes on batch size"},
          {"name": "Organization categorization", "type": "string", "description": "Category of the developing organization (e.g. Industry, Academia)"},
          {"name": "Foundation model", "type": "boolean", "description": "Whether this is a foundation model"},
          {"name": "Training compute lower bound", "type": "number", "description": "Lower bound estimate of training compute in FLOP"},
          {"name": "Training compute upper bound", "type": "number", "description": "Upper bound estimate of training compute in FLOP"},
          {"name": "Training chip-hours", "type": "number", "description": "Total chip-hours used for training"},
          {"name": "Training code accessibility", "type": "string", "description": "Accessibility of training code"},
          {"name": "Accessibility notes", "type": "string", "description": "Notes on accessibility of model or code"},
          {"name": "Organization categorization (from Organization)", "type": "string", "description": "Organization category derived from organization field"},
          {"name": "Possibly over 1e23 FLOP", "type": "boolean", "description": "Whether training compute may exceed 1e23 FLOP"},
          {"name": "Training compute cost (2023 USD)", "type": "number", "description": "Estimated training compute cost in 2023 US dollars"},
          {"name": "Training dataset size", "type": "number", "description": "Size of the training dataset (alternative field)"},
          {"name": "Sparsity", "type": "number", "description": "Model sparsity ratio"},
          {"name": "Utilization notes", "type": "string", "description": "Notes on hardware utilization"},
          {"name": "Estimated over 1e25 FLOP", "type": "boolean", "description": "Whether training compute is estimated to exceed 1e25 FLOP"},
          {"name": "Power per GPU", "type": "number", "description": "Power draw per GPU unit in watts"},
          {"name": "Cluster total TDP", "type": "number", "description": "Total thermal design power of the training cluster in watts"},
          {"name": "Base model compute", "type": "number", "description": "Training compute of the base model in FLOP"},
          {"name": "Total compute - (base + finetune)", "type": "number", "description": "Total compute including base model and fine-tuning in FLOP"},
          {"name": "API prices", "type": "string", "description": "API pricing information for the model"},
          {"name": "Created", "type": "string", "description": "Timestamp when the record was created"},
          {"name": "Inference code accessibility", "type": "string", "description": "Accessibility of inference code"},
          {"name": "Numerical format", "type": "string", "description": "Numerical precision format used in training (e.g. FP16, BF16)"},
          {"name": "Model versions", "type": "string", "description": "Available versions of the model"},
          {"name": "Frontier model", "type": "boolean", "description": "Whether this model was a frontier model at the time of release"},
          {"name": "Training power draw (W)", "type": "number", "description": "Power consumption during training in watts"},
          {"name": "Benchmark evals", "type": "string", "description": "Benchmark evaluation results"},
          {"name": "FLOP/$", "type": "number", "description": "Training compute efficiency in FLOP per dollar"},
          {"name": "Hardware release date", "type": "date", "format": "any", "description": "Release date of the training hardware"},
          {"name": "Hardware age", "type": "number", "description": "Age of the training hardware in years at time of training"},
          {"name": "Hardware FP32", "type": "number", "description": "Hardware FP32 FLOP/s throughput"},
          {"name": "Hardware TF32", "type": "number", "description": "Hardware TF32 FLOP/s throughput"},
          {"name": "Hardware count", "type": "number", "description": "Number of hardware units in the training cluster"},
          {"name": "Hardware TF16", "type": "number", "description": "Hardware TF16 FLOP/s throughput"},
          {"name": "Hardware FP16", "type": "number", "description": "Hardware FP16 FLOP/s throughput"},
          {"name": "Assumed precision", "type": "string", "description": "Assumed numerical precision for compute estimates"},
          {"name": "Assumed hardware FLOP/s", "type": "number", "description": "Assumed hardware throughput in FLOP/s used for compute estimates"},
          {"name": "Hardware type", "type": "string", "description": "Type of hardware used (e.g. GPU, TPU)"},
          {"name": "Training compute estimation method", "type": "string", "description": "Method used to estimate training compute"},
          {"name": "Biological model safeguards", "type": "string", "description": "Safeguards related to biological model risks"},
          {"name": "BenchmarkHub-v1", "type": "string", "description": "BenchmarkHub v1 evaluation results"},
          {"name": "Hugging Face developer id", "type": "string", "description": "Hugging Face developer or organization identifier"},
          {"name": "Post-training compute (FLOP)", "type": "number", "description": "Compute used for post-training (RLHF, fine-tuning, etc.) in FLOP"},
          {"name": "Post-training compute notes", "type": "string", "description": "Notes on post-training compute estimate"},
          {"name": "Hardware maker", "type": "string", "description": "Manufacturer of the training hardware"},
          {"name": "benchmarks/models", "type": "string", "description": "Benchmark to model mapping data"},
          {"name": "Maybe over 1e25 FLOP", "type": "boolean", "description": "Whether training compute may exceed 1e25 FLOP"},
          {"name": "Updated dataset size", "type": "number", "description": "Updated or revised training dataset size"},
          {"name": "WT103 ppl", "type": "number", "description": "WikiText-103 perplexity score"},
          {"name": "WT2 ppl", "type": "number", "description": "WikiText-2 perplexity score"},
          {"name": "PTB ppl", "type": "number", "description": "Penn Treebank perplexity score"},
          {"name": "Distillation or synthetic data", "type": "string", "description": "Whether model was trained on distillation or synthetic data"},
          {"name": "Distillation or synthetic data compute", "type": "number", "description": "Compute used to generate distillation or synthetic training data in FLOP"},
          {"name": "Distillation or synthetic data compute notes", "type": "string", "description": "Notes on distillation or synthetic data compute"},
          {"name": "Knowledge cutoff", "type": "string", "description": "Training data knowledge cutoff date"},
          {"name": "Context window", "type": "number", "description": "Maximum context window size in tokens"},
          {"name": "Hardware utilization (HFU)", "type": "number", "description": "Hardware FLOP utilization (HFU) during training"},
          {"name": "Training compute cost (cloud)", "type": "number", "description": "Estimated training compute cost using cloud pricing in USD"},
          {"name": "Training compute cost (upfront)", "type": "number", "description": "Estimated training compute cost using upfront hardware pricing in USD"}
        ]
      }
    }
  ]
}