Current-tokenizer cost (Δ, worst tokenizer)

These are the committed bytes rendered as readable JSON. Expanding this audit detail does not change the measurement’s current status.

{
    "metric": "token_delta",
    "formula_version": 1,
    "construct": "multiply-the-quantity",
    "models": [
        "cl100k_base",
        "o200k_base",
        "p50k_base"
    ],
    "test_set": [
        {
            "form": "as-quantity",
            "english": "Rack B stores 3 times more crates than rack A.",
            "ainglish": "Rack B stores 3 times as many crates as rack A."
        },
        {
            "form": "as-quantity",
            "english": "Depot B dispatches 4 times more shipments than depot A.",
            "ainglish": "Depot B dispatches 4 times as many shipments as depot A."
        },
        {
            "form": "as-quantity",
            "english": "Service B reports 6 times more errors than service A.",
            "ainglish": "Service B reports 6 times as many errors as service A."
        },
        {
            "form": "as-quantity",
            "english": "Link B carries 2 times more traffic than link A.",
            "ainglish": "Link B carries 2 times as much traffic as link A."
        },
        {
            "form": "as-quantity",
            "english": "Team B resolves 5 times more cases than team A.",
            "ainglish": "Team B resolves 5 times as many cases as team A."
        },
        {
            "form": "as-quantity",
            "english": "Lens B collects 8 times more light than lens A.",
            "ainglish": "Lens B collects 8 times as much light as lens A."
        },
        {
            "form": "times-the-noun",
            "english": "Cache B has 3 times more capacity than cache A.",
            "ainglish": "Cache B has 3 times the capacity of cache A."
        },
        {
            "form": "times-the-noun",
            "english": "Engine B produces 4 times more torque than engine A.",
            "ainglish": "Engine B produces 4 times the torque of engine A."
        },
        {
            "form": "times-the-noun",
            "english": "Archive B contains 7 times more records than archive A.",
            "ainglish": "Archive B contains 7 times the records of archive A."
        },
        {
            "form": "times-the-noun",
            "english": "Battery B provides 2 times more duration than battery A.",
            "ainglish": "Battery B provides 2 times the duration of battery A."
        },
        {
            "form": "times-the-noun",
            "english": "Queue B has 5 times more depth than queue A.",
            "ainglish": "Queue B has 5 times the depth of queue A."
        },
        {
            "form": "times-the-noun",
            "english": "Map B offers 10 times more resolution than map A.",
            "ainglish": "Map B offers 10 times the resolution of map A."
        },
        {
            "form": "notation",
            "english": "Worker B emits 3 times more records than worker A.",
            "ainglish": "Worker B emits 3× the records of worker A."
        },
        {
            "form": "notation",
            "english": "Bus B supplies 4 times more bandwidth than bus A.",
            "ainglish": "Bus B supplies 4× the bandwidth of bus A."
        },
        {
            "form": "notation",
            "english": "Farm B yields 6 times more grain than farm A.",
            "ainglish": "Farm B yields 6× the grain of farm A."
        },
        {
            "form": "notation",
            "english": "Antenna B delivers 2 times more gain than antenna A.",
            "ainglish": "Antenna B delivers 2× the gain of antenna A."
        },
        {
            "form": "notation",
            "english": "Compressor B generates 5 times more pressure than compressor A.",
            "ainglish": "Compressor B generates 5× the pressure of compressor A."
        },
        {
            "form": "notation",
            "english": "Index B holds 9 times more entries than index A.",
            "ainglish": "Index B holds 9× the entries of index A."
        },
        {
            "form": "fractional-decrease",
            "english": "Channel B has 3 times less noise than channel A.",
            "ainglish": "Channel B has one third the noise of channel A."
        },
        {
            "form": "fractional-decrease",
            "english": "Image B uses 2 times fewer bytes than image A.",
            "ainglish": "Image B uses half as many bytes as image A."
        },
        {
            "form": "fractional-decrease",
            "english": "Process B has 5 times lower latency than process A.",
            "ainglish": "Process B has one fifth the latency of process A."
        },
        {
            "form": "fractional-decrease",
            "english": "Build B needs 4 times less memory than build A.",
            "ainglish": "Build B needs a quarter of the memory of build A."
        },
        {
            "form": "fractional-decrease",
            "english": "Sample B contains 10 times less contamination than sample A.",
            "ainglish": "Sample B contains one tenth the contamination of sample A."
        },
        {
            "form": "fractional-decrease",
            "english": "Route B experiences 2 times fewer losses than route A.",
            "ainglish": "Route B experiences half as many losses as route A."
        }
    ],
    "seed": "none — deterministic tokenizer counts",
    "population": "24 fresh ratio-preserving repairs, balanced six each across as-quantity, times-the-noun, multiplication notation, and fractional-decrease forms",
    "selection": "All pairs were authored before tokenizer exposure and checked against every served prior test_set. Each refused comparative and conformant repair names the same subject, baseline, multiplier, quantity and intended ratio.",
    "method": "Under tiktoken 0.14.0, compute tokens(ainglish)-tokens(english) per pair on cl100k_base, o200k_base and p50k_base. Average 24 items equally per tokenizer and report the maximum tokenizer mean as least-favourable token_delta; value_lo/value_hi are the member span. Form means are diagnostic.",
    "estimand": {
        "population": "the 24 frozen different-input refused/conformant pairs",
        "aggregation": "equal-item mean per tokenizer; headline is maximum tokenizer mean",
        "comparator": "refused multiplicative comparative expressing the same intended ratio",
        "comparator_class": "refused_comparative"
    },
    "environment": {
        "tiktoken": "0.14.0",
        "python": "3.12.3"
    },
    "replicates_hash": "b6e9f938b17c27e8385a4bd4e22ad8c94e28ab6c1839722870d4dee9a64d0c5e",
    "freeze": "The API retains canonical manifest bytes before this process imports tiktoken or observes a count."
}