"""Chinese Classical Bench — leaderboard Space.

Single source of truth = the committed `leaderboard.md` (produced by the
GitHub repo's scripts/aggregate.py). This Space renders it verbatim so the
public leaderboard can never drift from the repo again. The interactive
table/chart are recomputed from the same results/*.json the markdown was
built from, so they always agree.

Primary metric per task matches scripts/aggregate.py: translate/char-gloss
use the Opus LLM judge (chrF under-rates synonymous paraphrase); chrF is the
labelled reproducible floor in the markdown's transparency table.
"""

import json
from pathlib import Path

import gradio as gr
import pandas as pd

HERE = Path(__file__).parent
RESULTS_DIR = HERE / "results"
LEADERBOARD_MD = HERE / "leaderboard.md"

PRIMARY = {
    "translate":    ("judge_norm",  "translate (Judge)"),
    "punctuate":    ("punct_f1",    "punctuate (Punct F1)"),
    "char-gloss":   ("judge_norm",  "char-gloss (Judge)"),
    "idiom-source": ("book_em",     "idiom-source (Book EM)"),
    "fill-in":      ("exact_match", "fill-in (Exact)"),
    "compress":     ("efficiency",  "compress (Eff)"),
}
TASK_ORDER = list(PRIMARY)


def load_leaderboard() -> pd.DataFrame:
    rows = []
    for fp in sorted(RESULTS_DIR.glob("*.json")):
        if fp.name.startswith("_"):
            continue
        d = json.loads(fp.read_text(encoding="utf-8"))
        row = {"Model": d.get("model", fp.stem)}
        vals, missing = [], []
        for t in TASK_ORDER:
            key, label = PRIMARY[t]
            v = d.get("tasks", {}).get(t, {}).get("summary", {}).get(key)
            if isinstance(v, (int, float)):
                row[label] = round(v, 3)
                vals.append(v)
            else:
                row[label] = None
                missing.append(t)
        cp = (d.get("tasks", {}).get("punctuate", {})
              .get("summary", {}).get("char_preserved"))
        row["Preserve"] = round(cp, 3) if isinstance(cp, (int, float)) else None
        row["Avg"] = round(sum(vals) / len(vals), 3) if vals else None
        row["_note"] = ("⚠ missing " + ",".join(missing)) if missing else ""
        rows.append(row)
    df = pd.DataFrame(rows)
    df = df.sort_values("Avg", ascending=False,
                        na_position="last").reset_index(drop=True)
    df.insert(0, "#", range(1, len(df) + 1))
    return df


DF = load_leaderboard()
TOP = DF.iloc[0]["Model"] if len(DF) else "—"
N_MODELS = len(DF)
LB_MD = (LEADERBOARD_MD.read_text(encoding="utf-8")
         if LEADERBOARD_MD.exists() else "_leaderboard.md missing_")

INTRO = f"""
# 🏛️ Chinese Classical Bench — Leaderboard

中国古典语言能力评测基准 — **6 个任务 × 100 题 = 600 道**（古译今 / 断句加标点 /
字义解释 / 典故出处 / 单字填空 / 今译古压缩）。回答一个具体问题：
**国产开源 LLM 在中国古典文献理解上谁更强？**

当前榜单：**{N_MODELS} 个模型**，按 primary 平均分榜首 **`{TOP}`**。

- 📦 评测集 (HF): [`gujilab/chinese-classical-bench`](https://huggingface.co/datasets/gujilab/chinese-classical-bench)
- 🧩 配套语料 (CC0): [`gujilab/chinese-classical-corpus`](https://huggingface.co/datasets/gujilab/chinese-classical-corpus)
- 💻 评测代码 / 完整方法与发现: [github.com/gujilab/chinese-classical-bench](https://github.com/gujilab/chinese-classical-bench)
  （见 `docs/findings.md` — 心理测量学审计、污染探针、judge 提升）

> **方法变更（v1.6+）**：`translate` / `char-gloss` 的 headline 已从 chrF
> 升级为 **Claude Opus 4.7 LLM judge**（chrF 系统性低估同义改写）；chrF 作为
> 可复现下限保留在下方 *Transparency* 表。新增 *Canonicity-stratified* 段：
> 按来源典籍知名度分层，暴露模型对"背过名篇"的依赖（`idiom-source` ρ=0.68）。
"""

with gr.Blocks(title="Chinese Classical Bench Leaderboard",
                theme=gr.themes.Soft()) as demo:
    gr.Markdown(INTRO)
    gr.Markdown("### Primary leaderboard（可排序）")
    gr.Dataframe(value=DF, interactive=False, wrap=True)
    gr.BarPlot(value=DF[["Model", "Avg"]].dropna(),
               x="Model", y="Avg", title="Primary 平均分 (Avg)")
    gr.Markdown("---\n以下为仓库 `leaderboard.md` 原文"
                "（含 canonicity-stratified 与 transparency 两表，"
                "与仓库严格一致）：\n")
    gr.Markdown(LB_MD)

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