src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

from src.about import Tasks


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


@dataclass
class AutoEvalColumn:
    model_type_symbol = ColumnContent("model_type_symbol", "str", True, never_hidden=True)
    model = ColumnContent("model", "markdown", True, never_hidden=True)
    average = ColumnContent("average", "number", True)
    anli = ColumnContent("ANLI", "number", True)
    logiqa = ColumnContent("LogiQA", "number", True)
    model_type = ColumnContent("model_type", "str", False)
    architecture = ColumnContent("architecture", "str", False)
    weight_type = ColumnContent("weight_type", "str", False, True)
    precision = ColumnContent("precision", "str", False)
    license = ColumnContent("license", "str", False)
    params = ColumnContent("#Params (B)", "number", False)
    likes = ColumnContent("Hub ❤️", "number", False)
    still_on_hub = ColumnContent("Available on the hub", "bool", False)
    revision = ColumnContent("Model sha", "str", False, False)


## 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")
    float32 = ModelDetails("float32")
    # 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 ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["float32"]:
            return Precision.float32
        # 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]

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]