src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

from src.display.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
    dummy: bool = False


## Leaderboard columns
auto_eval_column_dict = []
"""
# Init
auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
# Scores
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "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(["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)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)])
"""


auto_eval_column_dict.append(["eval_name", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", True)])
auto_eval_column_dict.append(["hf_model_id", ColumnContent, ColumnContent("Model URL", "str", False)])
auto_eval_column_dict.append(["aggregate_score", ColumnContent, ColumnContent("Aggregate Score", "number", True)])
auto_eval_column_dict.append(["grammar_avg", ColumnContent, ColumnContent("Grammar (Avg.)", "number", True)])
auto_eval_column_dict.append(["knowledge_avg", ColumnContent, ColumnContent("Knowledge (Avg.)", "number", True)])
auto_eval_column_dict.append(["reasoning_avg", ColumnContent, ColumnContent("Reasoning (Avg.)", "number", True)])
auto_eval_column_dict.append(["math_avg", ColumnContent, ColumnContent("Math (Avg.)", "number", True)])
auto_eval_column_dict.append(["classification_avg", ColumnContent, ColumnContent("Classification (Avg.)", "number", True)])
auto_eval_column_dict.append(["agree_cs", ColumnContent, ColumnContent("AGREE", "number", True)])
auto_eval_column_dict.append(["anli_cs", ColumnContent, ColumnContent("ANLI", "number", True)])
auto_eval_column_dict.append(["arc_challenge_cs", ColumnContent, ColumnContent("ARC-Challenge", "number", True)])
auto_eval_column_dict.append(["arc_easy_cs", ColumnContent, ColumnContent("ARC-Easy", "number", True)])
auto_eval_column_dict.append(["belebele_cs", ColumnContent, ColumnContent("Belebele", "number", True)])
auto_eval_column_dict.append(["ctkfacts_cs", ColumnContent, ColumnContent("CTKFacts", "number", True)])
auto_eval_column_dict.append(["czechnews_cs", ColumnContent, ColumnContent("Czech News", "number", True)])
auto_eval_column_dict.append(["fb_comments_cs", ColumnContent, ColumnContent("Facebook Comments", "number", True)])
auto_eval_column_dict.append(["gsm8k_cs", ColumnContent, ColumnContent("GSM8K", "number", True)])
auto_eval_column_dict.append(["klokanek_cs", ColumnContent, ColumnContent("Klokanek", "number", True)])
auto_eval_column_dict.append(["mall_reviews_cs", ColumnContent, ColumnContent("Mall Reviews", "number", True)])
auto_eval_column_dict.append(["mmlu_cs", ColumnContent, ColumnContent("MMLU", "number", True)])
auto_eval_column_dict.append(["sqad_cs", ColumnContent, ColumnContent("SQAD", "number", True)])
auto_eval_column_dict.append(["subjectivity_cs", ColumnContent, ColumnContent("Subjectivity", "number", True)])
auto_eval_column_dict.append(["truthfulqa_cs", ColumnContent, ColumnContent("TruthfulQA", "number", True)])
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent(" ", "str", True, dummy=True)])  # The dataframe does not display the last column - BUG in gradio?



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


HEADER_MAP = {
    "eval_name": "Model",
    "precision": "Precision",
    "hf_model_id": "Model URL",
    "knowledge_avg": "Knowledge (Avg.)",
    "agree_cs": "AGREE",
    "anli_cs": "ANLI",
    "arc_challenge_cs": "ARC-Challenge",
    "arc_easy_cs": "ARC-Easy",
    "belebele_cs": "Belebele",
    "ctkfacts_cs": "CTKFacts",
    "czechnews_cs": "Czech News",
    "fb_comments_cs": "Facebook Comments",
    "gsm8k_cs": "GSM8K",
    "klokanek_cs": "Klokanek",
    "mall_reviews_cs": "Mall Reviews",
    "mmlu_cs": "MMLU",
    "sqad_cs": "SQAD",
    "subjectivity_cs": "Subjectivity",
    "truthfulqa_cs": "TruthfulQA",
    "dummy": " ",
    "aggregate_score": "Aggregate Score",
}


## 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):
    other = ModelDetails("other")
    float64 = ModelDetails("float64")
    float32 = ModelDetails("float32")
    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.float64", "torch.double" ,"float64"]:
            return Precision.float64
        if precision in ["torch.float32", "torch.float" ,"float32"]:
            return Precision.tfloat32
        if precision in ["torch.float16", "torch.half", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["8bit", "int8"]:
            return Precision.qt_8bit
        if precision in ["4bit", "int4"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.other


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]

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

BENCHMARK_COLS = [HEADER_MAP[t.value.col_name] for t in Tasks]
BENCHMARK_COL_IDS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "?": pd.Interval(-1, 0, closed="right"),
    "~1.5": pd.Interval(0, 2, closed="right"),
    "~3": pd.Interval(2, 4, closed="right"),
    "~7": pd.Interval(4, 9, closed="right"),
    "~13": pd.Interval(9, 20, closed="right"),
    "~35": pd.Interval(20, 45, closed="right"),
    "~60": pd.Interval(45, 70, closed="right"),
    "70+": pd.Interval(70, 10000, closed="right"),
}