src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd


def fields(raw_class):
    return [
        v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"
    ]


@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str


class Tasks(Enum):
    arc = Task("arc_challenge", "acc_norm", "ARC")
    hellaswag = Task("hellaswag", "acc_norm", "HellaSwag")
    mmlu = Task("mmlu", "acc", "MMLU")
    truthfulqa = Task("truthfulqa_mc", "mc2", "TruthfulQA")
    # winogrande = Task("winogrande", "acc_norm", "Winogrande")
    # gsm8k = Task("gsm8k", "acc_norm", "GSM8k")
    # commongen_v2 = Task("commongen_v2", "acc_norm", "CommonGen V2")
    # eqBench = Task("eq_bench", "acc_norm", "EQ Bench")
    # instFollow = Task("inst_follow", "acc_norm", "InstFollow")
    # harmlessness = Task("harmlessness", "acc_norm", "Harmlessness")
    # helpfulness = Task("helpfulness", "acc_norm", "Helpfulness")


class Ranks(Enum):
    daily = Task("daily", "daily", "Daily Rank")
    quarterly = Task("quarterly", "quarterly", "Quarterly Rank")


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False


auto_eval_column_dict = []
# Init
auto_eval_column_dict.append(
    [
        "model_type_symbol",
        ColumnContent,
        ColumnContent("T", "str", True, never_hidden=True),
    ]
)
auto_eval_column_dict.append(
    [
        "model",
        ColumnContent,
        ColumnContent("Model", "markdown", True, never_hidden=True),
    ]
)
# Ranks
auto_eval_column_dict.append(
    ["daily", ColumnContent, ColumnContent("Daily Rank", "number", True)]
)
auto_eval_column_dict.append(
    ["quarterly", ColumnContent, ColumnContent("Quarterly Rank", "number", True)]
)
# Scores
auto_eval_column_dict.append(
    ["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)]
)
for task in Tasks:
    auto_eval_column_dict.append(
        [task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)]
    )
# Model information
auto_eval_column_dict.append(
    ["model_type", ColumnContent, ColumnContent("Type", "str", False)]
)
auto_eval_column_dict.append(
    ["architecture", ColumnContent, ColumnContent("Architecture", "str", False)]
)
auto_eval_column_dict.append(
    ["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)]
)
auto_eval_column_dict.append(
    ["precision", ColumnContent, ColumnContent("Precision", "str", False)]
)
auto_eval_column_dict.append(
    ["merged", ColumnContent, ColumnContent("Merged", "bool", False)]
)
auto_eval_column_dict.append(
    ["license", ColumnContent, ColumnContent("Hub License", "str", False)]
)
auto_eval_column_dict.append(
    ["params", ColumnContent, ColumnContent("#Params (B)", "number", False)]
)
auto_eval_column_dict.append(
    ["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)]
)
auto_eval_column_dict.append(
    [
        "still_on_hub",
        ColumnContent,
        ColumnContent("Available on the hub", "bool", False),
    ]
)
auto_eval_column_dict.append(
    ["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)]
)
auto_eval_column_dict.append(
    ["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, False)]
)
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(
    [
        "dummy",
        ColumnContent,
        ColumnContent("model_name_for_query", "str", False, dummy=True),
    ]
)

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)


@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    private = ColumnContent("private", "bool", True)
    precision = ColumnContent("precision", "str", True)
    weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


# Define the human baselines
human_baseline_row = {
    AutoEvalColumn.model.name: "<p>Human performance</p>",
}


@dataclass
class ModelDetails:
    name: str
    symbol: str = ""  # emoji, only for the model type


class ModelType(Enum):
    PT = ModelDetails(name="pretrained", symbol="🟢")
    # FT = ModelDetails(name="fine-tuned", symbol="🔶")
    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
    RL = ModelDetails(name="RL-tuned", symbol="🟦")
    Unknown = ModelDetails(name="", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        # if "fine-tuned" in type or "🔶" in type:
        #     return ModelType.FT
        if "pretrained" in type or "🟢" in type:
            return ModelType.PT
        if "RL-tuned" in type or "🟦" in type:
            return ModelType.RL
        if "instruction-tuned" in type or "⭕" in type:
            return ModelType.IFT
        return ModelType.Unknown


class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")


class Precision(Enum):
    float16 = ModelDetails("float16")
    # bfloat16 = ModelDetails("bfloat16")
    # qt_8bit = ModelDetails("8bit")
    # qt_4bit = ModelDetails("4bit")
    # qt_GPTQ = ModelDetails("GPTQ")
    Unknown = ModelDetails("?")

    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["8bit"]:
            return Precision.qt_8bit
        if precision in ["4bit"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.Unknown


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
COLS_LITE = [
    c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden
]
TYPES_LITE = [
    c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden
]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "Unknown": pd.Interval(-1, 0, closed="right"),
    "0~3B": pd.Interval(0, 3, closed="right"),
    "3~7B": pd.Interval(3, 7.3, closed="right"),
    "7~13B": pd.Interval(7.3, 13, closed="right"),
    "13~35B": pd.Interval(13, 35, closed="right"),
    "35~60B": pd.Interval(35, 60, closed="right"),
    "60B+": pd.Interval(60, 10000, closed="right"),
}