src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

from src.about import Tasks, TaskType


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_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
auto_eval_column_dict.append(["n_shot", ColumnContent, ColumnContent("N-Shot", "number", False)])
auto_eval_column_dict.append(["prompt_version", ColumnContent, ColumnContent("Version", "str", False)])
#Scores
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average (All) ⬆️", "number", True)])
for task_type in TaskType:
    auto_eval_column_dict.append([task_type.value.name, ColumnContent, ColumnContent(f"Average ({task_type.value.display_name}) {task_type.value.symbol}", "number", True)])
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", task.value.is_primary_metric)])
# Model information
auto_eval_column_dict.append(["model_training", ColumnContent, ColumnContent("Type", "str", False)])
auto_eval_column_dict.append(["maltese_training", ColumnContent, ColumnContent("Maltese Training", "str", False)])
auto_eval_column_dict.append(["language_count", ColumnContent, ColumnContent("#Languages", "number", 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(["license", ColumnContent, ColumnContent("Hub License", "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)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)

## 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)
    precision = ColumnContent("precision", "str", True)
    n_shot = ColumnContent("n_shot", "int", True)
    prompt_version = ColumnContent("prompt_version", "str", True)
    seed = ColumnContent("seed", "int", True)
    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 ModelTraining(Enum):
    PT = ModelDetails(name="pre-trained", symbol="PT")
    FT = ModelDetails(name="fine-tuned", symbol="FT")
    IT = ModelDetails(name="instruction-tuned", symbol="IT")
    NK = ModelDetails(name="unknown", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        type = type or ""
        if "PT" in type:
            return ModelTraining.PT
        if "FT" in type:
            return ModelTraining.FT
        if "IT" in type:
            return ModelTraining.IT
        return ModelTraining.NK


class MalteseTraining(Enum):
    NO = ModelDetails(name="none", symbol="NO")
    PT = ModelDetails(name="pre-training", symbol="PT")
    FT = ModelDetails(name="fine-tuning", symbol="FT")
    IT = ModelDetails(name="instruction-tuning", symbol="IT")
    NK = ModelDetails(name="unknown", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        type = type or ""
        if "NO" in type:
            return MalteseTraining.NO
        if "PT" in type:
            return MalteseTraining.PT
        if "FT" in type:
            return MalteseTraining.FT
        if "IT" in type:
            return MalteseTraining.IT
        return MalteseTraining.NK

class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")

class Precision(Enum):
    float32 = ModelDetails("float32")
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    Unknown = ModelDetails("?")

    def from_str(precision):
        if precision in ["torch.float32", "float32"]:
            return Precision.float32
        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]

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]