from dataclasses import dataclass, make_dataclass, field from enum import Enum import pandas as pd from src.about import Tasks, Domains def fields(raw_class): return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"] # 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 ## Leaderboard columns auto_eval_column_dict = [] # # Init # auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)]) # auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)]) # # new columns # for domain in Domains: # auto_eval_column_dict.append([domain.name, ColumnContent, ColumnContent(domain.value.col_name, "number", True)]) # auto_eval_column_dict.append(["organization", ColumnContent, ColumnContent("Organization", "str", False)]) # auto_eval_column_dict.append(["knowledge_cutoff", ColumnContent, ColumnContent("Knowledge cutoff", "str", False)]) # for task in Tasks: # auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)]) # auto_eval_column_dict.append(["model_type_symbol", ColumnContent, field(default_factory=lambda: ColumnContent("T", "str", True, never_hidden=True))]) # #Scores # auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "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(["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)]) # Init auto_eval_column_dict.append(["model", ColumnContent, field(default_factory=lambda: ColumnContent("Model", "markdown", True, never_hidden=True))]) auto_eval_column_dict.append(["license", ColumnContent, field(default_factory=lambda: ColumnContent("License", "str", False))]) # new columns for domain in Domains: auto_eval_column_dict.append([domain.name, ColumnContent, field(default_factory=lambda: ColumnContent(domain.value.col_name, "number", True))]) auto_eval_column_dict.append(["organization", ColumnContent, field(default_factory=lambda: ColumnContent("Organization", "str", False))]) auto_eval_column_dict.append(["knowledge_cutoff", ColumnContent, field(default_factory=lambda: ColumnContent("Knowledge cutoff", "str", False))]) auto_eval_column_dict.append(["score", ColumnContent, field(default_factory=lambda: ColumnContent("Average Score", "number", True))]) auto_eval_column_dict.append(["score_sd", ColumnContent, field(default_factory=lambda: ColumnContent("Score SD", "number", True))]) auto_eval_column_dict.append(["rank", ColumnContent, field(default_factory=lambda: ColumnContent("Rank", "number", True))]) # fine-grained dimensions auto_eval_column_dict.append(["score_overall", ColumnContent, field(default_factory=lambda: ColumnContent("Score (MT-Bench)", "number", True))]) auto_eval_column_dict.append(["sd_overall", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev(MT-Bench)", "number", True))]) auto_eval_column_dict.append(["rank_overall", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (MT-Bench)", "number", True))]) auto_eval_column_dict.append(["score_math_algebra", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Math Algebra)", "number", True))]) auto_eval_column_dict.append(["sd_math_algebra", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Math Algebra)", "number", True))]) auto_eval_column_dict.append(["rank_math_algebra", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Math Algebra)", "number", True))]) auto_eval_column_dict.append(["score_math_geometry", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Math Geometry)", "number", True))]) auto_eval_column_dict.append(["sd_math_geometry", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Math Geometry)", "number", True))]) auto_eval_column_dict.append(["rank_math_geometry", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Math Geometry)", "number", True))]) auto_eval_column_dict.append(["score_math_probability", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Math Probability)", "number", True))]) auto_eval_column_dict.append(["sd_math_probability", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Math Probability)", "number", True))]) auto_eval_column_dict.append(["rank_math_probability", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Math Probability)", "number", True))]) auto_eval_column_dict.append(["score_reason_logical", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Logical Reasoning)", "number", True))]) auto_eval_column_dict.append(["sd_reason_logical", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Logical Reasoning)", "number", True))]) auto_eval_column_dict.append(["rank_reason_logical", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Logical Reasoning)", "number", True))]) auto_eval_column_dict.append(["score_reason_social", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Social Reasoning)", "number", True))]) auto_eval_column_dict.append(["sd_reason_social", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Social Reasoning)", "number", True))]) auto_eval_column_dict.append(["rank_reason_social", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Social Reasoning)", "number", True))]) auto_eval_column_dict.append(["score_chemistry", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Chemistry)", "number", True))]) auto_eval_column_dict.append(["sd_chemistry", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Chemistry)", "number", True))]) auto_eval_column_dict.append(["rank_chemistry", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Chemistry)", "number", True))]) auto_eval_column_dict.append(["score_physics", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Physics)", "number", True))]) auto_eval_column_dict.append(["sd_physics", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Physics)", "number", True))]) auto_eval_column_dict.append(["rank_physics", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Physics)", "number", True))]) auto_eval_column_dict.append(["score_biology", ColumnContent, field(default_factory=lambda: ColumnContent("Score (Biology)", "number", True))]) auto_eval_column_dict.append(["sd_biology", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (Biology)", "number", True))]) auto_eval_column_dict.append(["rank_biology", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (Biology)", "number", True))]) auto_eval_column_dict.append(["score_cpp", ColumnContent, field(default_factory=lambda: ColumnContent("Score (C++)", "number", True))]) auto_eval_column_dict.append(["sd_cpp", ColumnContent, field(default_factory=lambda: ColumnContent("Std dev (C++)", "number", True))]) auto_eval_column_dict.append(["rank_cpp", ColumnContent, field(default_factory=lambda: ColumnContent("Rank (C++)", "number", True))]) for task in Tasks: auto_eval_column_dict.append([task.name, ColumnContent, field(default_factory=lambda: ColumnContent(task.value.col_name, "number", True))]) auto_eval_column_dict.append(["model_type_symbol", ColumnContent, field(default_factory=lambda: ColumnContent("T", "str", True, never_hidden=True))]) #Scores auto_eval_column_dict.append(["average", ColumnContent, field(default_factory=lambda: ColumnContent("Average ⬆️", "number", True))]) # Model information auto_eval_column_dict.append(["model_type", ColumnContent, field(default_factory=lambda: ColumnContent("Type", "str", False))]) auto_eval_column_dict.append(["architecture", ColumnContent, field(default_factory=lambda: ColumnContent("Architecture", "str", False))]) auto_eval_column_dict.append(["weight_type", ColumnContent, field(default_factory=lambda: ColumnContent("Weight type", "str", False, True))]) auto_eval_column_dict.append(["precision", ColumnContent, field(default_factory=lambda: ColumnContent("Precision", "str", False))]) auto_eval_column_dict.append(["params", ColumnContent, field(default_factory=lambda: ColumnContent("#Params (B)", "number", False))]) auto_eval_column_dict.append(["likes", ColumnContent, field(default_factory=lambda: ColumnContent("Hub ❤️", "number", False))]) auto_eval_column_dict.append(["still_on_hub", ColumnContent, field(default_factory=lambda: ColumnContent("Available on the hub", "bool", False))]) auto_eval_column_dict.append(["revision", ColumnContent, field(default_factory=lambda: ColumnContent("Model sha", "str", False, False))]) # We use make dataclass to dynamically fill the scores from Tasks AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True) AutoEvalColumn = AutoEvalColumn() # print all attributes of AutoEvalColumn # print(AutoEvalColumn.__annotations__.keys()) # preint precision attribute # print(AutoEvalColumn.precision) ## For the queue columns in the submission tab @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) ## All the model information that we might need @dataclass class ModelDetails: name: str display_name: str = "" symbol: str = "" # emoji 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") Unknown = ModelDetails("?") def from_str(precision): if precision in ["torch.float16", "float16"]: return Precision.float16 if precision in ["torch.bfloat16", "bfloat16"]: return Precision.bfloat16 return Precision.Unknown # Column selection COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden] # print(COLS) 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]