src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd

from src.about import Tasks,Quotas

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(["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(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
auto_eval_column_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
# 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)])

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


## Leaderboard columns
auto_eval_column_quota_dict = []
# Init
auto_eval_column_quota_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_quota_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
#Scores
auto_eval_column_quota_dict.append(["average_quota", ColumnContent, ColumnContent("AverageSampled ⬆️", "number", True)])
for task in Quotas:
    auto_eval_column_quota_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
# Model information
auto_eval_column_quota_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
auto_eval_column_quota_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
auto_eval_column_quota_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
auto_eval_column_quota_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
auto_eval_column_quota_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
auto_eval_column_quota_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
auto_eval_column_quota_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_quota_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumnQuota = make_dataclass("AutoEvalColumnQuota", auto_eval_column_quota_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)
    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 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 "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) if not c.hidden]
TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]

QUOTACOLS = [c.name for c in fields(AutoEvalColumnQuota) if not c.hidden]
QUOTATYPES = [c.type for c in fields(AutoEvalColumnQuota) 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]
BENCHMARK_QUOTACOLS = [t.value.col_name for t in Quotas]

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"),
}

# Define the baselines
#baseline_row = {
#    AutoEvalColumn.model.name: "<p>Baseline</p>",
#    AutoEvalColumn.revision.name: "N/A",
#    AutoEvalColumn.precision.name: None,
#    AutoEvalColumn.average.name: 92.75,
#    #AutoEvalColumn.merged.name: False,
#    AutoEvalColumn.CMMMU.name: 100,
#    AutoEvalColumn.MMMU.name: 100,
#    AutoEvalColumn.MMMU_Pro_standard.name: 100,
#    AutoEvalColumn.MMMU_Pro_vision.name: 100,
#    AutoEvalColumn.MathVision.name: 100,
#    AutoEvalColumn.CII_Bench.name: 100,
#    AutoEvalColumn.Blink.name: 100,
#    AutoEvalColumn.CharXiv.name: 100,
#    AutoEvalColumn.MathVerse.name: 100,
#    AutoEvalColumn.MmvetV2.name: 100,
#    AutoEvalColumn.Ocrlite.name: 100,
#    AutoEvalColumn.OcrliteZh.name: 100,
#    AutoEvalColumn.dummy.name: "baseline",
#    AutoEvalColumn.model_type.name: "",
#    AutoEvalColumn.flagged.name: False,
#}
#
## Define the human baselines
#human_baseline_row = {
#    AutoEvalColumn.model.name: "<p>Human performance</p>",
#    AutoEvalColumn.revision.name: "N/A",
#    AutoEvalColumn.precision.name: None,
#    AutoEvalColumn.average.name: 92.75,
#    #AutoEvalColumn.merged.name: False,
#    AutoEvalColumn.CMMMU.name: 100,
#    AutoEvalColumn.MMMU.name: 100,
#    AutoEvalColumn.MMMU_Pro_standard.name: 100,
#    AutoEvalColumn.MMMU_Pro_vision.name: 100,
#    AutoEvalColumn.MathVision.name: 100,
#    AutoEvalColumn.CII_Bench.name: 100,
#    AutoEvalColumn.Blink.name: 100,
#    AutoEvalColumn.CharXiv.name: 100,
#    AutoEvalColumn.MathVerse.name: 100,
#    AutoEvalColumn.MmvetV2.name: 100,
#    AutoEvalColumn.Ocrlite.name: 100,
#    AutoEvalColumn.OcrliteZh.name: 100,
#    AutoEvalColumn.dummy.name: "human_baseline",
#    AutoEvalColumn.model_type.name: "",
#    AutoEvalColumn.flagged.name: False,
#}