src/display/utils.py

from dataclasses import dataclass, make_dataclass

import pandas as pd


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", ColumnContent, ColumnContent("Model Name", "markdown", True, never_hidden=True)]
)
auto_eval_column_dict.append(["model_provider", ColumnContent, ColumnContent("LLM Provider", "markdown", True)])
auto_eval_column_dict.append(["use_case_name", ColumnContent, ColumnContent("Use Case Name", "markdown", True)])
auto_eval_column_dict.append(["use_case_type", ColumnContent, ColumnContent("Use Case Type", "markdown", False)])
auto_eval_column_dict.append(["accuracy_method", ColumnContent, ColumnContent("Accuracy Method", "markdown", False)])
# Accuracy metrics
auto_eval_column_dict.append(["accuracy_metric_average", ColumnContent, ColumnContent("Accuracy", "markdown", True)])
auto_eval_column_dict.append(
    [
        "accuracy_metric_instruction_following",
        ColumnContent,
        ColumnContent("Instruction Following", "markdown", True),
    ]
)
auto_eval_column_dict.append(
    ["accuracy_metric_completeness", ColumnContent, ColumnContent("Completeness", "markdown", True)]
)
auto_eval_column_dict.append(
    ["accuracy_metric_conciseness", ColumnContent, ColumnContent("Conciseness", "markdown", True)]
)
auto_eval_column_dict.append(
    ["accuracy_metric_factuality", ColumnContent, ColumnContent("Factuality", "markdown", True)]
)
# Speed (Latency) & Cost metrics
auto_eval_column_dict.append(["latency", ColumnContent, ColumnContent("Response Time (Sec)", "markdown", True)])
auto_eval_column_dict.append(
    ["mean_output_tokens", ColumnContent, ColumnContent("Mean Output Tokens", "markdown", True)]
)
auto_eval_column_dict.append(["cost_band", ColumnContent, ColumnContent("Cost Band", "markdown", True)])
# Trust & Safety metrics
auto_eval_column_dict.append(["ts", ColumnContent, ColumnContent("Trust & Safety", "markdown", True)])
auto_eval_column_dict.append(["safety", ColumnContent, ColumnContent("Safety", "markdown", False)])
auto_eval_column_dict.append(["privacy", ColumnContent, ColumnContent("Privacy", "markdown", False)])
auto_eval_column_dict.append(["truthfulness", ColumnContent, ColumnContent("Truthfulness", "markdown", False)])
auto_eval_column_dict.append(["crm_bias", ColumnContent, ColumnContent("CRM Bias", "markdown", False)])
# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)


# Speed (Latency) & Cost metrics
cost_eval_column_dict = []
# Init
cost_eval_column_dict.append(
    ["model", ColumnContent, ColumnContent("Model Name", "markdown", True, never_hidden=True)]
)
cost_eval_column_dict.append(["model_provider", ColumnContent, ColumnContent("LLM Provider", "markdown", True)])
cost_eval_column_dict.append(
    ["use_case_flavor", ColumnContent, ColumnContent("Cost and Speed: Flavor", "markdown", True)]
)
cost_eval_column_dict.append(["latency", ColumnContent, ColumnContent("Response Time (Sec)", "markdown", True)])
cost_eval_column_dict.append(
    ["mean_output_tokens", ColumnContent, ColumnContent("Mean Output Tokens", "markdown", True)]
)
cost_eval_column_dict.append(["cost_band", ColumnContent, ColumnContent("Cost Band", "markdown", True)])
CostEvalColumn = make_dataclass("CostEvalColumn", cost_eval_column_dict, frozen=True)

# Trust & Safety metrics
ts_eval_column_dict = []
# Init
ts_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model Name", "markdown", True, never_hidden=True)])
# ts_eval_column_dict.append(["model_provider", ColumnContent, ColumnContent("LLM Provider", "markdown", True)])
ts_eval_column_dict.append(["ts", ColumnContent, ColumnContent("Trust & Safety", "markdown", True)])
ts_eval_column_dict.append(["safety", ColumnContent, ColumnContent("Safety", "markdown", False)])
ts_eval_column_dict.append(["privacy", ColumnContent, ColumnContent("Privacy", "markdown", False)])
ts_eval_column_dict.append(["truthfulness", ColumnContent, ColumnContent("Truthfulness", "markdown", False)])
ts_eval_column_dict.append(["crm_bias", ColumnContent, ColumnContent("CRM Bias", "markdown", False)])
# ts_eval_column_dict.append(["bias_no_ci", ColumnContent, ColumnContent("Bias No CI", "markdown", True)])
TSEvalColumn = make_dataclass("TSEvalColumn", ts_eval_column_dict, frozen=True)

# 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]

COST_COLS = [c.name for c in fields(CostEvalColumn) if not c.hidden]
COST_TYPES = [c.type for c in fields(CostEvalColumn) if not c.hidden]

TS_COLS = [c.name for c in fields(TSEvalColumn) if not c.hidden]
TS_TYPES = [c.type for c in fields(TSEvalColumn) if not c.hidden]

# BENCHMARK_COLS = [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"),
}