src/streamlit_app.py

"""Streamlit app to display the NPU Arabic leaderboard."""

from __future__ import annotations

import json
from datetime import datetime
from pathlib import Path
from typing import List, Sequence

import streamlit as st

_DATA_PATH = Path("../leaderboard.json")
_COLUMNS: Sequence[str] = (
    "model_name",
    "avg_tps",
    "iwslt2017-en-ar_sacrebleu",
    "mlqa_ar_ar_f1",
    "xquad_ar_f1",
    "timestamp",
)


def _load_rows() -> List[dict]:
    if not _DATA_PATH.exists():
        return []
    try:
        raw = json.loads(_DATA_PATH.read_text(encoding="utf-8"))
    except json.JSONDecodeError:
        return []

    if isinstance(raw, dict):
        data = [raw]
    elif isinstance(raw, list):
        data = [item for item in raw if isinstance(item, dict)]
    else:
        data = []

    # Filter to desired columns and sort newest-first.
    filtered: List[dict] = []
    for row in data:
        compact = {key: row.get(key) for key in _COLUMNS}
        filtered.append(compact)

    def _sort_key(item: dict) -> tuple:
        stamp = item.get("timestamp")
        try:
            return (datetime.fromisoformat(str(stamp)),)
        except Exception:
            return (datetime.min,)

    filtered.sort(key=_sort_key, reverse=True)
    return filtered


st.set_page_config(page_title="Intel NPU Arabic Leaderboard", layout="wide")
st.title("Intel® NPU Arabic Leaderboard")

rows = _load_rows()
if not rows:
    st.info("No evaluations uploaded yet. Trigger a run to populate the leaderboard.")
else:
    st.write(
        "Latest evaluation per model. Add new results by emailing the evaluation endpoint "
        "or running the CLI with the Hugging Face publishing flags."
    )
    st.dataframe(rows, column_config={col: st.column_config.Column(col) for col in _COLUMNS})

st.caption("Data auto-synced from leaderboard.json produced by the evaluation pipeline.")