intigration/SecuritySupervisor

1from dataclasses import dataclass, make_dataclass2from enum import Enum3 4import pandas as pd5 6from src.about import Tasks7 8def fields(raw_class):9    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]10 11 12# These classes are for user facing column names,13# to avoid having to change them all around the code14# when a modif is needed15@dataclass16class ColumnContent:17    name: str18    type: str19    displayed_by_default: bool20    hidden: bool = False21    never_hidden: bool = False22 23## Leaderboard columns24auto_eval_column_dict = []25# Init26auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])27auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])28#Scores29auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])30for task in Tasks:31    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])32# Model information33auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])34auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])35auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])36auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])37auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])38auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])39auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])40auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)])41auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])42 43# We use make dataclass to dynamically fill the scores from Tasks44AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)45 46## For the queue columns in the submission tab47@dataclass(frozen=True)48class EvalQueueColumn:  # Queue column49    model = ColumnContent("model", "markdown", True)50    revision = ColumnContent("revision", "str", True)51    private = ColumnContent("private", "bool", True)52    precision = ColumnContent("precision", "str", True)53    weight_type = ColumnContent("weight_type", "str", "Original")54    status = ColumnContent("status", "str", True)55 56## All the model information that we might need57@dataclass58class ModelDetails:59    name: str60    display_name: str = ""61    symbol: str = "" # emoji62 63 64class ModelType(Enum):65    PT = ModelDetails(name="pretrained", symbol="🟢")66    FT = ModelDetails(name="fine-tuned", symbol="🔶")67    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")68    RL = ModelDetails(name="RL-tuned", symbol="🟦")69    Unknown = ModelDetails(name="", symbol="?")70 71    def to_str(self, separator=" "):72        return f"{self.value.symbol}{separator}{self.value.name}"73 74    @staticmethod75    def from_str(type):76        if "fine-tuned" in type or "🔶" in type:77            return ModelType.FT78        if "pretrained" in type or "🟢" in type:79            return ModelType.PT80        if "RL-tuned" in type or "🟦" in type:81            return ModelType.RL82        if "instruction-tuned" in type or "⭕" in type:83            return ModelType.IFT84        return ModelType.Unknown85 86class WeightType(Enum):87    Adapter = ModelDetails("Adapter")88    Original = ModelDetails("Original")89    Delta = ModelDetails("Delta")90 91class Precision(Enum):92    float16 = ModelDetails("float16")93    bfloat16 = ModelDetails("bfloat16")94    float32 = ModelDetails("float32")95    #qt_8bit = ModelDetails("8bit")96    #qt_4bit = ModelDetails("4bit")97    #qt_GPTQ = ModelDetails("GPTQ")98    Unknown = ModelDetails("?")99 100    def from_str(precision):101        if precision in ["torch.float16", "float16"]:102            return Precision.float16103        if precision in ["torch.bfloat16", "bfloat16"]:104            return Precision.bfloat16105        if precision in ["float32"]:106            return Precision.float32107        #if precision in ["8bit"]:108        #    return Precision.qt_8bit109        #if precision in ["4bit"]:110        #    return Precision.qt_4bit111        #if precision in ["GPTQ", "None"]:112        #    return Precision.qt_GPTQ113        return Precision.Unknown114 115# Column selection116COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]117TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]118COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]119TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]120 121EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]122EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]123 124BENCHMARK_COLS = [t.value.col_name for t in Tasks]125 126NUMERIC_INTERVALS = {127    "?": pd.Interval(-1, 0, closed="right"),128    "~1.5": pd.Interval(0, 2, closed="right"),129    "~3": pd.Interval(2, 4, closed="right"),130    "~7": pd.Interval(4, 9, closed="right"),131    "~13": pd.Interval(9, 20, closed="right"),132    "~35": pd.Interval(20, 45, closed="right"),133    "~60": pd.Interval(45, 70, closed="right"),134    "70+": pd.Interval(70, 10000, closed="right"),135}136