src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum
import json
import logging
from datetime import datetime
import pandas as pd


# Configure logging
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")


def parse_datetime(datetime_str):
    formats = [
        "%Y-%m-%dT%H-%M-%S.%f",  # Format with dashes
        "%Y-%m-%dT%H:%M:%S.%f",  # Standard format with colons
        "%Y-%m-%dT%H %M %S.%f",  # Spaces as separator
    ]

    for fmt in formats:
        try:
            return datetime.strptime(datetime_str, fmt)
        except ValueError:
            continue
    # in rare cases set unix start time for files with incorrect time (legacy files)
    logging.error(f"No valid date format found for: {datetime_str}")
    return datetime(1970, 1, 1)


def load_json_data(file_path):
    """Safely load JSON data from a file."""
    try:
        with open(file_path, "r") as file:
            return json.load(file)
    except json.JSONDecodeError:
        print(f"Error reading JSON from {file_path}")
        return None  # Or raise an exception


def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str


class Tasks(Enum):
    math = Task("RussianMath", "score", "math_score")
    physics = Task("RussianPhysics", "score", "physics_score")
    combined = Task("Combined", "score", "score")


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass(frozen=True)
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False


auto_eval_column_dict = []
# Init
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("model", "markdown", True, never_hidden=True)])
# Scores
auto_eval_column_dict.append(["score", ColumnContent, ColumnContent("score", "number", True)])
for task in Tasks:
    if task != Tasks.combined:  # Combined score уже добавлен выше
        auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])

# Model information
auto_eval_column_dict.append(["total_tokens", ColumnContent, ColumnContent("total_tokens", "number", False)])
auto_eval_column_dict.append(["evaluation_time", ColumnContent, ColumnContent("evaluation_time", "number", False)])
auto_eval_column_dict.append(["system_prompt", ColumnContent, ColumnContent("system_prompt", "str", False)])

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


@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)


baseline_row = {
    AutoEvalColumn.model.name: "<p>Baseline</p>",
    AutoEvalColumn.score.name: 0.1,
    AutoEvalColumn.math.name: 0.1,
    AutoEvalColumn.physics.name: 0.1,
    AutoEvalColumn.total_tokens.name: 0,
    AutoEvalColumn.evaluation_time.name: 0,
    AutoEvalColumn.system_prompt.name: "Вы - полезный помощник по математике и физике. Ответьте на русском языке.",
}

# Define the human baselines
human_baseline_row = {
    AutoEvalColumn.model.name: "<p>Human performance</p>",
    AutoEvalColumn.score.name: 0.9,
    AutoEvalColumn.math.name: 0.9,
    AutoEvalColumn.physics.name: 0.9,
    AutoEvalColumn.total_tokens.name: 0,
    AutoEvalColumn.evaluation_time.name: 0,
    AutoEvalColumn.system_prompt.name: "Вы - полезный помощник по математике и физике. Ответьте на русском языке.",
}


@dataclass
class ModelDetails:
    name: str
    symbol: str = ""  # emoji, only for the model type


class ModelType(Enum):
    PT = ModelDetails(name="pretrained", symbol="🟢")
    CPT = ModelDetails(name="continuously pretrained", symbol="🟩")
    FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶")
    chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬")
    merges = ModelDetails(name="base merges and moerges", 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 "continously pretrained" in type or "🟩" in type:
            return ModelType.CPT
        if "pretrained" in type or "🟢" in type:
            return ModelType.PT
        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
            return ModelType.chat
        if "merge" in type or "🤝" in type:
            return ModelType.merges
        return ModelType.Unknown


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


class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    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 ["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)]
TYPES = [c.type for c in fields(AutoEvalColumn)]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

NUMERIC_INTERVALS = {
    "?": pd.Interval(-1, 0, closed="right"),
    "~0.1": pd.Interval(0, 0.2, closed="right"),
    "~0.3": pd.Interval(0.2, 0.4, closed="right"),
    "~0.5": pd.Interval(0.4, 0.6, closed="right"),
    "~0.7": pd.Interval(0.6, 0.8, closed="right"),
    "0.8+": pd.Interval(0.8, 1.0, closed="right"),
}