logikon/open_cot_leaderboard

1from dataclasses import dataclass, make_dataclass2from enum import Enum3from typing import Any4 5import pandas as pd # type: ignore6 7from src.display.about import Tasks8 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@dataclass17class ColumnContent:18    name: str19    type: str20    displayed_by_default: bool21    hidden: bool = False22    never_hidden: bool = False23    dummy: bool = False24 25## Leaderboard columns26auto_eval_column_dict: list[tuple[str, type, Any]] = []27# Init28auto_eval_column_dict.append(("model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)))29auto_eval_column_dict.append(("model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)))30# Scores31auto_eval_column_dict.append(("average", ColumnContent, ColumnContent("Average ⬆️", "number", True)))32for task in Tasks:33    auto_eval_column_dict.append((task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)))34# Dashboard35auto_eval_column_dict.append(("dashboard_link", ColumnContent, ColumnContent("Dashboard", "markdown", True)))36# Model information37auto_eval_column_dict.append(("model_type", ColumnContent, ColumnContent("Type", "str", False)))38auto_eval_column_dict.append(("architecture", ColumnContent, ColumnContent("Architecture", "str", False)))39auto_eval_column_dict.append(("weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)))40auto_eval_column_dict.append(("precision", ColumnContent, ColumnContent("Precision", "str", False)))41auto_eval_column_dict.append(("license", ColumnContent, ColumnContent("Hub License", "str", False)))42auto_eval_column_dict.append(("params", ColumnContent, ColumnContent("#Params (B)", "number", False)))43auto_eval_column_dict.append(("likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)))44auto_eval_column_dict.append(("still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)))45auto_eval_column_dict.append(("revision", ColumnContent, ColumnContent("Model sha", "str", False, False)))46# Dummy column for the search bar (hidden by the custom CSS)47auto_eval_column_dict.append(("dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)))48 49# We use make dataclass to dynamically fill the scores from Tasks50AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)51 52## For the queue columns in the submission tab53@dataclass(frozen=True)54class EvalQueueColumn:  # Queue column55    model = ColumnContent("model", "markdown", True)56    revision = ColumnContent("revision", "str", True)57    private = ColumnContent("private", "bool", True)58    precision = ColumnContent("precision", "str", True)59    weight_type = ColumnContent("weight_type", "str", True)60    status = ColumnContent("status", "str", True)61 62## All the model information that we might need63@dataclass64class ModelDetails:65    name: str66    display_name: str = ""67    symbol: str = "" # emoji68 69 70class ModelType(Enum):71    PT = ModelDetails(name="pretrained", symbol="🟢")72    FT = ModelDetails(name="fine-tuned", symbol="🔶")73    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")74    RL = ModelDetails(name="RL-tuned", symbol="🟦")75    Unknown = ModelDetails(name="", symbol="?")76 77    def to_str(self, separator=" "):78        return f"{self.value.symbol}{separator}{self.value.name}"79 80    @staticmethod81    def from_str(type):82        if "fine-tuned" in type or "🔶" in type:83            return ModelType.FT84        if "pretrained" in type or "🟢" in type:85            return ModelType.PT86        if "RL-tuned" in type or "🟦" in type:87            return ModelType.RL88        if "instruction-tuned" in type or "⭕" in type:89            return ModelType.IFT90        return ModelType.Unknown91 92class WeightType(Enum):93    Adapter = ModelDetails("Adapter")94    Original = ModelDetails("Original")95    Delta = ModelDetails("Delta")96 97class Precision(Enum):98    float16 = ModelDetails("float16")99    bfloat16 = ModelDetails("bfloat16")100    qt_8bit = ModelDetails("8bit")101    qt_4bit = ModelDetails("4bit")102    qt_GPTQ = ModelDetails("GPTQ")103    Unknown = ModelDetails("?")104 105    def from_str(precision):106        if precision in ["torch.float16", "float16"]:107            return Precision.float16108        if precision in ["torch.bfloat16", "bfloat16"]:109            return Precision.bfloat16110        if precision in ["8bit"]:111            return Precision.qt_8bit112        if precision in ["4bit"]:113            return Precision.qt_4bit114        if precision in ["GPTQ", "None"]:115            return Precision.qt_GPTQ116        return Precision.Unknown117 118# Column selection119COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]120TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]121COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]122TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]123 124EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]125EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]126 127BENCHMARK_COLS = [t.value.col_name for t in Tasks]128 129NUMERIC_INTERVALS = {130    "?": pd.Interval(-1, 0, closed="right"),131    "~1.5": pd.Interval(0, 2, closed="right"),132    "~3": pd.Interval(2, 4, closed="right"),133    "~7": pd.Interval(4, 9, closed="right"),134    "~13": pd.Interval(9, 20, closed="right"),135    "~35": pd.Interval(20, 45, closed="right"),136    "~60": pd.Interval(45, 70, closed="right"),137    "70+": pd.Interval(70, 10000, closed="right"),138}139