MLRS/MELABench

1from dataclasses import dataclass, make_dataclass2from enum import Enum3 4import pandas as pd5 6from src.about import Tasks, TaskType7 8 9def fields(raw_class):10    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]11 12 13# These classes are for user facing column names,14# to avoid having to change them all around the code15# when a modif is needed16@dataclass(frozen=True)17class ColumnContent:18    name: str19    type: str20    displayed_by_default: bool21    hidden: bool = False22    never_hidden: bool = False23    hidden_in_fewshot: bool = False24 25## Leaderboard columns26auto_eval_column_dict = []27# Init28auto_eval_column_dict.append(["model_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])29auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])30auto_eval_column_dict.append(["n_shot", ColumnContent, ColumnContent("N-Shot", "number", True)])31auto_eval_column_dict.append(["prompt_version", ColumnContent, ColumnContent("Version", "str", True)])32#Scores33auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average (All) ⬆️", "number", True)])34for task_type in TaskType:35    auto_eval_column_dict.append([task_type.value.name, ColumnContent, ColumnContent(f"Average ({task_type.value.display_name}) {task_type.value.symbol}", "number", True)])36for task in Tasks:37    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", task.value.is_primary_metric, hidden_in_fewshot=task.value.zero_shot_only)])38# Model information39auto_eval_column_dict.append(["model_training", ColumnContent, ColumnContent("Type", "str", False)])40auto_eval_column_dict.append(["maltese_training", ColumnContent, ColumnContent("Maltese Training", "str", False)])41auto_eval_column_dict.append(["language_count", ColumnContent, ColumnContent("#Languages", "number", False)])42auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])43auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])44auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])45auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])46auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])47auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])48auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)])49auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model SHA", "str", False, False)])50 51# We use make dataclass to dynamically fill the scores from Tasks52AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)53 54## For the queue columns in the submission tab55@dataclass(frozen=True)56class EvalQueueColumn:  # Queue column57    model = ColumnContent("model", "markdown", True)58    revision = ColumnContent("revision", "str", True)59    precision = ColumnContent("precision", "str", True)60    n_shot = ColumnContent("n_shot", "int", True)61    prompt_version = ColumnContent("prompt_version", "str", True)62    seed = ColumnContent("seed", "int", True)63    status = ColumnContent("status", "str", True)64 65## All the model information that we might need66@dataclass67class ModelDetails:68    name: str69    display_name: str = ""70    symbol: str = "" # emoji71 72 73class ModelTraining(Enum):74    PT = ModelDetails(name="pre-trained", symbol="PT")75    FT = ModelDetails(name="fine-tuned", symbol="FT")76    IT = ModelDetails(name="instruction-tuned", symbol="IT")77    NK = ModelDetails(name="unknown", symbol="?")78 79    def to_str(self, separator=" "):80        return f"{self.value.symbol}{separator}{self.value.name}"81 82    @staticmethod83    def from_str(type):84        type = type or ""85        if "PT" in type:86            return ModelTraining.PT87        if "FT" in type:88            return ModelTraining.FT89        if "IT" in type:90            return ModelTraining.IT91        return ModelTraining.NK92 93 94class MalteseTraining(Enum):95    NO = ModelDetails(name="none", symbol="NO")96    PT = ModelDetails(name="pre-training", symbol="PT")97    FT = ModelDetails(name="fine-tuning", symbol="FT")98    IT = ModelDetails(name="instruction-tuning", symbol="IT")99    NK = ModelDetails(name="unknown", symbol="?")100 101    def to_str(self, separator=" "):102        return f"{self.value.symbol}{separator}{self.value.name}"103 104    @staticmethod105    def from_str(type):106        type = type or ""107        if "NO" in type:108            return MalteseTraining.NO109        if "PT" in type:110            return MalteseTraining.PT111        if "FT" in type:112            return MalteseTraining.FT113        if "IT" in type:114            return MalteseTraining.IT115        return MalteseTraining.NK116 117class WeightType(Enum):118    Adapter = ModelDetails("Adapter")119    Original = ModelDetails("Original")120    Delta = ModelDetails("Delta")121 122class Precision(Enum):123    float32 = ModelDetails("float32")124    float16 = ModelDetails("float16")125    bfloat16 = ModelDetails("bfloat16")126    Unknown = ModelDetails("?")127 128    def from_str(precision):129        if precision in ["torch.float32", "float32"]:130            return Precision.float32131        if precision in ["torch.float16", "float16"]:132            return Precision.float16133        if precision in ["torch.bfloat16", "bfloat16"]:134            return Precision.bfloat16135        return Precision.Unknown136 137# Column selection138COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]139 140EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]141EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]142 143BENCHMARK_COLS = [t.value.col_name for t in Tasks]144