src/about.py

from __future__ import annotations

TITLE = "MOOD Bench Leaderboard"

METRIC_KEY = "tpr@fpr0.01"
METRIC_LABEL = "TPR @ FPR=1%"

INTRO_MARKDOWN = f"""
# MOOD Bench Leaderboard

Multi-domain out-of-distribution safety detection on the
[`mood-bench`](https://huggingface.co/datasets/mood-bench/mood-bench) test split.

Each row is a **detection method**. For every out-of-distribution unsafe domain we 
report the true-positive rate at a fixed **1% false-positive rate** against the 
pooled in-distribution *safe* conversations. The **Overall** column is the unweighted 
mean across OOD-unsafe domains.

Use the **Columns** picker below to show the per-domain breakdown, or the filter 
controls to narrow the table down. Sorted by Overall {METRIC_LABEL}.
"""

GROUPS: list[str] = [
    "overall",
    "id",
    "controlling",
    "insecure-code",
    "scheming",
    "jailbroken",
    "sycophantic",
    "function-calling-missing",
    "function-calling-inappropriate",
    "swahili",
]

GROUP_LABELS: dict[str, str] = {
    "overall": "Overall",
    "id": "ID",
    "controlling": "Controlling",
    "insecure-code": "Insecure Code",
    "scheming": "Scheming",
    "jailbroken": "Jailbroken",
    "sycophantic": "Sycophantic",
    "function-calling-missing": "FC Missing",
    "function-calling-inappropriate": "FC Inapprop.",
    "swahili": "Swahili",
}

META_COLUMNS: list[str] = ["Method", "Model", "Submitted By", "Submitted At"]


def metric_column(group: str) -> str:
    """Stable column header for the flattened DataFrame."""
    return GROUP_LABELS[group]


ALL_METRIC_COLUMNS: list[str] = [metric_column(g) for g in GROUPS]

DEFAULT_VISIBLE_COLUMNS: list[str] = [
    "Method",
    "Model",
    metric_column("overall"),
    metric_column("id"),
]

PRIMARY_SORT_COLUMN = metric_column("overall")