app.py

"""ConStellaration Progress Tracker.

Read a single results.jsonl from the tracker dataset, render it in the same
leaderboard pattern as the official Space. No submit / evaluate paths.
"""
import json
import os
from pathlib import Path

import gradio as gr
import pandas as pd
from gradio_leaderboard import Leaderboard
from huggingface_hub import hf_hub_download

RESULTS_REPO = "phanerozoic/constellaration-tracker-results"
RESULTS_FILE = "results.jsonl"


def make_user_link(name: str) -> str:
    return (
        f'<a target="_blank" href="https://huggingface.co/{name}" '
        f'style="color: var(--link-text-color); text-decoration: underline;'
        f'text-decoration-style: dotted;">{name}</a>'
    )


def make_bundle_link(path: str) -> str:
    if not path:
        return ""
    url = f"https://huggingface.co/datasets/{RESULTS_REPO}/blob/main/{path}"
    return (
        f'<a target="_blank" href="{url}" '
        f'style="color: var(--link-text-color); text-decoration: underline;'
        f'text-decoration-style: dotted;">bundle</a>'
    )


def load_results_df() -> pd.DataFrame:
    try:
        local = hf_hub_download(
            repo_id=RESULTS_REPO,
            repo_type="dataset",
            filename=RESULTS_FILE,
            force_download=True,
        )
        records = [json.loads(line) for line in Path(local).read_text().splitlines() if line.strip()]
    except Exception as e:
        records = [{"error": str(e)}]
    if not records or "error" in records[0]:
        return pd.DataFrame(columns=["submission time", "problem type", "user", "score", "bundle size", "bundle"])
    df = pd.DataFrame(records)
    df.rename(columns={
        "submission_time": "submission time",
        "problem_type": "problem type",
        "bundle_size": "bundle size",
        "bundle_filename": "bundle",
    }, inplace=True)
    if "user" in df.columns:
        df["user"] = df["user"].apply(make_user_link).astype(str)
    if "bundle" in df.columns:
        df["bundle"] = df["bundle"].apply(make_bundle_link).astype(str)
    cols = [c for c in ["submission time", "problem type", "user", "score", "bundle size", "bundle"] if c in df.columns]
    df = df[cols]
    if "score" in df.columns:
        df = df.sort_values("score", ascending=False)
    return df


def gradio_interface() -> gr.Blocks:
    with gr.Blocks() as demo:
        gr.Markdown("## ConStellaration Progress Tracker")
        gr.Markdown(
            "Unofficial fallback view of locally-evaluated ConStellaration "
            "benchmark submissions. The "
            "[official leaderboard](https://huggingface.co/spaces/proxima-fusion/constellaration-bench) "
            "remains the authoritative source for ranked submissions; this Space exists "
            "so progress remains visible during periods when the official Space is down "
            "or its result-write step is delayed."
        )
        with gr.Tabs(elem_classes="tab-buttons"):
            with gr.TabItem("🛰️ Tracker"):
                Leaderboard(
                    value=load_results_df(),
                    datatype=["date", "str", "html", "number", "number", "html"],
                    select_columns=["submission time", "problem type", "user", "score", "bundle size", "bundle"],
                    search_columns=["submission time", "score", "user", "problem type"],
                    filter_columns=["problem type"],
                    every=120,
                    render=True,
                )
                gr.Markdown(
                    "Scores are computed by running the standard `constellaration.problems` "
                    "evaluation locally. For `geometrical` and `simple_to_build`, scores are "
                    "bounded between 0.0 and 1.0. For `mhd_stable`, the score is unbounded "
                    "hypervolume of the feasible Pareto front in (-L_grad_B, aspect_ratio) "
                    "space with reference point (1.0, 20.0)."
                )

            with gr.TabItem("❔ About"):
                gr.Markdown(
                    """
                    ### Purpose

                    The official [ConStellaration leaderboard](https://huggingface.co/spaces/proxima-fusion/constellaration-bench)
                    relies on a server-side scoring pipeline. When that pipeline is down or
                    delayed, submitted boundaries do not appear on the ranking even though
                    they have been uploaded. This Space mirrors what the official table
                    would show using local evaluations of the same scoring code.

                    ### How scoring works

                    Every entry shown here is produced by running
                    `constellaration.problems.<problem>().evaluate(boundary)` on the same
                    boundary file that was (or will be) submitted to the official Space.
                    The scoring code is the public reference implementation shipped in the
                    [`constellaration`](https://github.com/proximafusion/constellaration)
                    package; no surrogate or approximation is used.

                    ### What this Space is not

                    - Not an official Proxima Fusion artifact.
                    - Not a substitute for the official leaderboard when it is up.
                    - Not a submission interface; new entries are added by writing to the
                      backing dataset `phanerozoic/constellaration-tracker-results`.

                    ### Dataset

                    [`phanerozoic/constellaration-tracker-results`](https://huggingface.co/datasets/phanerozoic/constellaration-tracker-results)
                    is a public dataset with a single `results.jsonl` file. Each line is one
                    locally-evaluated record.
                    """
                )

    return demo


if __name__ == "__main__":
    gradio_interface().launch()