AINativeBench / data /processed /RQ3 /plot_total_tokens_violin.py
王子睿
Update processed data
c4e7970 7 months ago
35.4 kB
#!/usr/bin/env python3

import csv
import os
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Tuple

import numpy as np

try:
    import matplotlib.pyplot as plt
    from matplotlib.ticker import FuncFormatter

    HAS_MATPLOTLIB = True
except ModuleNotFoundError:
    plt = None  # type: ignore[assignment]
    FuncFormatter = None  # type: ignore[assignment]
    HAS_MATPLOTLIB = False

if HAS_MATPLOTLIB:
    plt.rcParams["font.family"] = "Times New Roman"

MODEL_ORDER: List[str] = [
    "GPT-5",
    "GPT-4o-mini",
    "DeepSeek-V3-1",
    "DeepSeek-R1",
    "Gemini-2.5-flash",
    "Gemini-2.5-flash-nothinking",
    "Qwen3-235b",
]

MODEL_LABELS: Dict[str, str] = {
    "GPT-5": "GPT-5",
    "GPT-4o-mini": "GPT-4o-mini",
    "DeepSeek-V3-1": "DeepSeek-V3.1",
    "DeepSeek-R1": "DeepSeek-R1",
    "Gemini-2.5-flash": "Gemini-2.5",
    "Gemini-2.5-flash-nothinking": "Gemini-2.5-NT",
    "Qwen3-235b": "Qwen3-235b",
}


ARCH_ORDER: List[str] = [
    "Unknown",
    "MCP",
    "A2A",
    "A2A_mix",
]

MODEL_COLORS: Dict[str, str] = {
    "GPT-5": "#1f77b4",
    "GPT-4o-mini": "#ff7f0e",
    "DeepSeek-V3-1": "#2ca02c",
    "DeepSeek-R1": "#d62728",
    "Gemini-2.5-flash": "#9467bd",
    "Gemini-2.5-flash-nothinking": "#8c564b",
    "Qwen3-235b": "#e377c2",
}

STATUS_ORDER: List[str] = [
    "success_no_retry",
    "success_with_retry",
    "failed",
]

STATUS_TITLES: Dict[str, str] = {
    "success_no_retry": "Pass (no retries)",
    "success_with_retry": "Pass (with retries)",
    "failed": "Failure",
}


def infer_project_dir(file_path: str) -> str:
    raw = (file_path or "").strip()
    if not raw:
        return ""
    try:
        p = Path(raw)
        return p.parents[2].name
    except Exception:
        return ""


def infer_architecture(project_dir: str) -> str:
    name = (project_dir or "").strip()
    if not name:
        return "Unknown"
    if name.endswith("-MCP"):
        return "MCP"
    if name.endswith("-H_A2A") or name.endswith("-H-A2A"):
        return "A2A_mix"
    if name.endswith("-A2A"):
        return "A2A"
    return "Unknown"


def infer_base_task(project_dir: str) -> str:
    name = (project_dir or "").strip()
    for suffix in ("-H_A2A", "-H-A2A", "-MCP", "-A2A"):
        if name.endswith(suffix):
            return name[: -len(suffix)]
    return name


def make_project_name(base_task: str, arch: str) -> str:
    task = (base_task or "").strip()
    if not task:
        return ""
    if arch == "Unknown":
        return task
    if arch == "MCP":
        return f"{task}-MCP"
    if arch == "A2A_mix":
        return f"{task}-H-A2A"
    if arch == "A2A":
        return f"{task}-A2A"
    return f"{task}-{arch}"


def infer_architecture_from_project_name(project_name: str) -> str:
    name = (project_name or "").strip()
    if not name:
        return "Unknown"
    if name.endswith("-MCP"):
        return "MCP"
    if name.endswith("-H-A2A") or name.endswith("-H_A2A"):
        return "A2A_mix"
    if name.endswith("-A2A"):
        return "A2A"
    return "Unknown"


def base_task_from_project_name(project_name: str) -> str:
    name = (project_name or "").strip()
    for suffix in ("-H-A2A", "-H_A2A", "-MCP", "-A2A"):
        if name.endswith(suffix):
            return name[: -len(suffix)]
    return name


def export_violin_input_summary(
    projects: List[Tuple[str, str, str]],
    project_data: Dict[str, Dict[str, Dict[str, List[float]]]],
    out_dir: Path,
) -> None:
    if (os.environ.get("EXPORT_VIOLIN_INPUT") or "").strip() not in {
        "1",
        "true",
        "True",
    }:
        return

    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / "violin_input_summary.csv"

    with out_path.open("w", encoding="utf-8", newline="") as f:
        writer = csv.writer(f)
        writer.writerow(
            [
                "project",
                "base_task",
                "architecture",
                "status_group",
                "model",
                "n",
                "mean",
                "median",
                "min",
                "max",
            ]
        )

        for _, _, name in projects:
            pdata = project_data.get(name, {})
            base_task = base_task_from_project_name(name)
            arch = infer_architecture_from_project_name(name)
            for status in STATUS_ORDER:
                by_model = pdata.get(status, {})
                for model in MODEL_ORDER:
                    vals = by_model.get(model, [])
                    if not vals:
                        continue
                    arr = np.asarray(vals, dtype=float)
                    writer.writerow(
                        [
                            name,
                            base_task,
                            arch,
                            status,
                            model,
                            int(arr.size),
                            float(np.mean(arr)),
                            float(np.median(arr)),
                            float(np.min(arr)),
                            float(np.max(arr)),
                        ]
                    )

    print(f"saved violin input summary: {out_path}")


def _nice_step(max_val: float, target_ticks: int = 6) -> float:
    if max_val <= 0:
        return 1.0
    raw = max_val / float(target_ticks)
    magnitude = 10 ** int(np.floor(np.log10(raw)))
    residual = raw / magnitude
    if residual <= 1:
        nice = 1
    elif residual <= 2:
        nice = 2
    elif residual <= 5:
        nice = 5
    else:
        nice = 10
    return nice * magnitude


def _token_formatter(x, pos):
    if x >= 1_000_000:
        return f"{x / 1_000_000:.1f}M"
    if x >= 1000:
        return f"{int(x // 1000)}k"
    return str(int(x))


def load_projects(
    details_csv: Path,
) -> Tuple[List[Tuple[str, str, str]], Dict[Tuple[str, str], str]]:
    present: Dict[str, set] = defaultdict(set)

    with details_csv.open("r", encoding="utf-8", newline="") as f:
        reader = csv.DictReader(f)
        for row in reader:
            project_dir = infer_project_dir(row.get("file_path") or "")
            if not project_dir:
                continue
            arch = infer_architecture(project_dir)
            base_task = infer_base_task(project_dir)
            if not base_task:
                continue

            model = (row.get("model") or "").strip()
            if model and model not in MODEL_ORDER:
                continue

            total_raw = row.get("total_tokens")
            if total_raw is None or total_raw == "":
                continue
            try:
                float(total_raw)
            except ValueError:
                continue

            present[base_task].add(arch)

    projects: List[Tuple[str, str, str]] = []
    project_map: Dict[Tuple[str, str], str] = {}
    for base_task in sorted(present.keys()):
        for arch in ARCH_ORDER:
            if arch not in present[base_task]:
                continue
            name = make_project_name(base_task, arch)
            project_map[(base_task, arch)] = name
            projects.append((base_task, arch, name))

    return projects, project_map


def classify_status(status_raw: str, with_retry_raw: str) -> str:
    s = (status_raw or "").strip().lower()
    w = (with_retry_raw or "").strip().lower()
    if s == "success" and w == "false":
        return "success_no_retry"
    if s == "success" and w == "true":
        return "success_with_retry"
    return "failed"


def load_total_tokens(
    details_csv: Path, project_map: Dict[Tuple[str, str], str]
) -> Dict[str, Dict[str, Dict[str, List[float]]]]:
    data: Dict[str, Dict[str, Dict[str, List[float]]]] = defaultdict(
        lambda: defaultdict(lambda: defaultdict(list))
    )

    with details_csv.open("r", encoding="utf-8", newline="") as f:
        reader = csv.DictReader(f)
        for row in reader:
            project_dir = infer_project_dir(row.get("file_path") or "")
            if not project_dir:
                continue
            task = infer_base_task(project_dir)
            arch = infer_architecture(project_dir)
            key = (task, arch)
            project_name = project_map.get(key) or make_project_name(task, arch)
            if not project_name:
                continue

            model = (row.get("model") or "").strip()
            if model not in MODEL_ORDER:
                continue

            status_group = classify_status(row.get("status"), row.get("with_retry"))

            total_raw = row.get("total_tokens")
            if total_raw is None or total_raw == "":
                continue
            try:
                total_val = float(total_raw)
            except ValueError:
                continue

            data[project_name][status_group][model].append(total_val)

    return data


def load_all_token_data(
    details_csv: Path, project_map: Dict[Tuple[str, str], str]
) -> Tuple[
    Dict[str, Dict[str, Dict[str, List[float]]]],
    Dict[str, Dict[str, Dict[str, List[float]]]],
    Dict[str, List[float]],
]:
    """
    Load token statistics with two views:
    - project_data: keyed by project name (task-architecture) -> status -> model -> list of totals
    - arch_model_data: keyed by task -> architecture -> model -> list of totals (across all statuses)
    - overall_status_values: keyed by status -> all token totals
    """
    project_data: Dict[str, Dict[str, Dict[str, List[float]]]] = defaultdict(
        lambda: defaultdict(lambda: defaultdict(list))
    )
    arch_model_data: Dict[str, Dict[str, Dict[str, List[float]]]] = defaultdict(
        lambda: defaultdict(lambda: defaultdict(list))
    )
    overall_status_values: Dict[str, List[float]] = defaultdict(list)

    with details_csv.open("r", encoding="utf-8", newline="") as f:
        reader = csv.DictReader(f)
        for row in reader:
            project_dir = infer_project_dir(row.get("file_path") or "")
            if not project_dir:
                continue

            task = infer_base_task(project_dir)
            arch = infer_architecture(project_dir)
            key = (task, arch)
            project_name = project_map.get(key) or make_project_name(task, arch)
            if not project_name:
                continue

            model = (row.get("model") or "").strip()
            if model not in MODEL_ORDER:
                continue

            status_group = classify_status(row.get("status"), row.get("with_retry"))

            total_raw = row.get("total_tokens")
            if total_raw is None or total_raw == "":
                continue
            try:
                total_val = float(total_raw)
            except ValueError:
                continue

            project_data[project_name][status_group][model].append(total_val)
            arch_model_data[task][arch][model].append(total_val)
            overall_status_values[status_group].append(total_val)

    return project_data, arch_model_data, overall_status_values


def summarize_values(values: List[float]) -> Dict[str, float]:
    if not values:
        return {}
    arr = np.asarray(values, dtype=float)
    return {
        "count": int(arr.size),
        "mean": float(np.mean(arr)),
    }


def _fmt_number(val: float) -> str:
    return f"{val:,.0f}"


def _fmt_mean(stats: Dict[str, float]) -> str:
    if not stats:
        return "-"
    return _fmt_number(stats.get("mean", 0.0))


def _safe_mean(values: List[float]) -> float:
    if not values:
        return None  # type: ignore[return-value]
    return float(np.mean(np.asarray(values, dtype=float)))


def _fmt_mean_val(val: float) -> str:
    if val is None:
        return "-"
    return _fmt_number(val)


def _fmt_delta(new_val: float, old_val: float) -> str:
    if new_val is None or old_val is None:
        return "-"
    diff = new_val - old_val
    pct_str = "n/a" if old_val == 0 else f"{(diff / old_val) * 100:.1f}%"
    sign = "+" if diff >= 0 else ""
    return f"{sign}{_fmt_number(diff)} ({pct_str})"


def generate_project_stats_md(
    projects: List[Tuple[str, str, str]],
    project_data: Dict[str, Dict[str, Dict[str, List[float]]]],
    overall_status_values: Dict[str, List[float]],
    out_dir: Path,
) -> None:
    def _base_task_name(task: str) -> str:
        suffixes = ("-H_A2A", "-H-A2A", "-MCP", "-A2A")
        for suffix in suffixes:
            if task.endswith(suffix):
                return task[: -len(suffix)]
        return task

    def build_series_data() -> Dict[str, Dict[str, Dict[str, List[float]]]]:
        series: Dict[str, Dict[str, Dict[str, List[float]]]] = defaultdict(
            lambda: defaultdict(lambda: defaultdict(list))
        )
        for task, arch, name in projects:
            base_task = _base_task_name(task)
            pdata = project_data.get(name, {})
            for status in STATUS_ORDER:
                by_model = pdata.get(status, {})
                for model, vals in by_model.items():
                    series[base_task][status][model].extend(vals)
        return series

    def append_status_table(
        lines: List[str],
        by_status: Dict[str, Dict[str, List[float]]],
        statuses: List[str],
    ) -> None:
        for status in statuses:
            by_model = by_status.get(status, {})
            if not by_model:
                continue
            lines.append("")
            lines.append(f"### {STATUS_TITLES.get(status, status)}")
            lines.append("")
            lines.append("| Model | n | Mean |")
            lines.append("| --- | --- | --- |")
            for model in MODEL_ORDER:
                vals = by_model.get(model, [])
                stats = summarize_values(vals)
                mean = _fmt_mean(stats)
                lines.append(
                    f"| {MODEL_LABELS.get(model, model)} | {stats.get('count', 0)} | {mean} |"
                )

    def append_status_comparison(
        lines: List[str],
        label: str,
        by_status: Dict[str, Dict[str, List[float]]],
        base_status: str,
        comp_status: str,
    ) -> None:
        lines.append("")
        lines.append(label)
        lines.append("")
        lines.append(
            f"| Model | {STATUS_TITLES[base_status]} mean | {STATUS_TITLES[comp_status]} mean | Δ vs {STATUS_TITLES[base_status]} |"
        )
        lines.append("| --- | --- | --- | --- |")
        for model in MODEL_ORDER:
            base_mean = _safe_mean(by_status.get(base_status, {}).get(model, []))
            comp_mean = _safe_mean(by_status.get(comp_status, {}).get(model, []))
            lines.append(
                "| "
                + " | ".join(
                    [
                        MODEL_LABELS.get(model, model),
                        _fmt_mean_val(base_mean),
                        _fmt_mean_val(comp_mean),
                        _fmt_delta(comp_mean, base_mean),
                    ]
                )
                + " |"
            )

    def build_report(
        title: str,
        base_status: str,
        comp_status: str,
        filename: str,
    ) -> None:
        statuses = [base_status, comp_status]
        series_data = build_series_data()
        lines: List[str] = []
        lines.append(f"# Project token statistics - {title}")
        lines.append("")
        lines.append("## Overall status token summary")
        lines.append("")
        lines.append("| Status | n | Mean |")
        lines.append("| --- | --- | --- |")
        for status in statuses:
            stats = summarize_values(overall_status_values.get(status, []))
            mean = _fmt_mean(stats)
            lines.append(
                f"| {STATUS_TITLES.get(status, status)} | {stats.get('count', 0)} | {mean} |"
            )

        for task, arch, name in projects:
            pdata = project_data.get(name, {})
            lines.append("")
            lines.append(f"## {name}")
            if not pdata:
                lines.append("")
                lines.append("> No token data found.")
                continue
            append_status_table(lines, pdata, statuses)
            append_status_comparison(
                lines,
                f"### {STATUS_TITLES[base_status]} vs {STATUS_TITLES[comp_status]} (mean, abs & %)",
                pdata,
                base_status,
                comp_status,
            )

        # Series-level aggregation by task prefix
        lines.append("")
        lines.append("## Series aggregates (by task prefix)")
        for task in sorted(series_data.keys()):
            sdata = series_data[task]
            lines.append("")
            lines.append(f"### {task} (aggregated across variants)")
            append_status_table(lines, sdata, statuses)
            append_status_comparison(
                lines,
                f"#### {STATUS_TITLES[base_status]} vs {STATUS_TITLES[comp_status]} (mean, abs & %)",
                sdata,
                base_status,
                comp_status,
            )

        out_path = out_dir / filename
        out_dir.mkdir(parents=True, exist_ok=True)
        out_path.write_text("\n".join(lines), encoding="utf-8")
        print(f"saved markdown: {out_path}")

    build_report(
        "Pass (no retry) vs Pass (with retry)",
        "success_no_retry",
        "success_with_retry",
        "project_token_stats_pass_vs_retry.md",
    )
    build_report(
        "Pass (no retry) vs Failure",
        "success_no_retry",
        "failed",
        "project_token_stats_pass_vs_failure.md",
    )


def generate_architecture_deltas_md(
    projects: List[Tuple[str, str, str]],
    arch_model_data: Dict[str, Dict[str, Dict[str, List[float]]]],
    out_dir: Path,
) -> None:
    def write_report(title: str, chain: List[str], filename: str) -> None:
        lines: List[str] = []
        lines.append("# Token shifts across architectures")
        lines.append("")
        lines.append(f"Series: **{title}**")
        lines.append("")
        lines.append(
            "Each table shows the absolute change (Δ) and the relative percentage change of the mean total tokens."
        )

        def render_pair(task_arch_data: Dict[str, Dict[str, List[float]]]) -> None:
            arch_display = {
                "Unknown": "Pure CrewAI",
                "MCP": "MCP",
                "A2A": "A2A",
                "A2A_mix": "H-A2A",
            }

            header = f"| Model | {arch_display[chain[0]]} | {arch_display[chain[1]]} | Δ {arch_display[chain[1]]}-{arch_display[chain[0]]} |"
            sep = "| --- | --- | --- | --- |"

            lines.append("")
            lines.append(header)
            lines.append(sep)

            for model in MODEL_ORDER:
                left_vals = task_arch_data.get(chain[0], {}).get(model, [])
                right_vals = task_arch_data.get(chain[1], {}).get(model, [])
                left_mean = float(np.mean(left_vals)) if left_vals else None
                right_mean = float(np.mean(right_vals)) if right_vals else None
                row = [
                    MODEL_LABELS.get(model, model),
                    "-" if left_mean is None else _fmt_number(left_mean),
                    "-" if right_mean is None else _fmt_number(right_mean),
                    _fmt_delta(right_mean, left_mean),
                ]
                lines.append("| " + " | ".join(row) + " |")

            lines.append("")
            lines.append("Project-level average (all models combined)")
            lines.append("")
            lines.append(
                f"| Metric | {arch_display[chain[0]]} | {arch_display[chain[1]]} | Δ {arch_display[chain[1]]}-{arch_display[chain[0]]} |"
            )
            lines.append("| --- | --- | --- | --- |")

            def _mean_all(arch: str) -> float:
                combined: List[float] = []
                for vals in task_arch_data.get(arch, {}).values():
                    combined.extend(vals)
                return float(np.mean(combined)) if combined else None

            left_all = _mean_all(chain[0])
            right_all = _mean_all(chain[1])
            lines.append(
                "| "
                + " | ".join(
                    [
                        "Avg tokens (all models)",
                        "-" if left_all is None else _fmt_number(left_all),
                        "-" if right_all is None else _fmt_number(right_all),
                        _fmt_delta(right_all, left_all),
                    ]
                )
                + " |"
            )

        task_set = {t for t, _, _ in projects}
        for task in sorted(task_set):
            task_arch_data = arch_model_data.get(task, {})
            arches = set(task_arch_data.keys())
            if not task_arch_data:
                continue
            if not set(chain).issubset(arches):
                continue

            lines.append("")
            lines.append(f"## {task}")
            render_pair(task_arch_data)

        out_path = out_dir / filename
        out_dir.mkdir(parents=True, exist_ok=True)
        out_path.write_text("\n".join(lines), encoding="utf-8")
        print(f"saved markdown: {out_path}")

    def write_a2a_to_h_a2a_report(filename: str) -> None:
        lines: List[str] = []
        lines.append("# Token shifts across architectures")
        lines.append("")
        lines.append("Series: **A2A → H-A2A**")
        lines.append("")
        lines.append(
            "Each table shows the absolute change (Δ) and the relative percentage change of the mean total tokens."
        )

        def render_pair(
            task_base: str,
            left_arch_data: Dict[str, Dict[str, List[float]]],
            right_arch_data: Dict[str, Dict[str, List[float]]],
        ) -> None:
            right_label = "H-A2A"
            header = f"| Model | A2A | {right_label} | Δ {right_label}-A2A |"
            sep = "| --- | --- | --- | --- |"
            lines.append("")
            lines.append(header)
            lines.append(sep)

            for model in MODEL_ORDER:
                left_vals = left_arch_data.get("A2A", {}).get(model, [])
                right_vals: List[float] = []
                for arch_vals in right_arch_data.values():
                    right_vals.extend(arch_vals.get(model, []))
                left_mean = float(np.mean(left_vals)) if left_vals else None
                right_mean = float(np.mean(right_vals)) if right_vals else None
                row = [
                    MODEL_LABELS.get(model, model),
                    "-" if left_mean is None else _fmt_number(left_mean),
                    "-" if right_mean is None else _fmt_number(right_mean),
                    _fmt_delta(right_mean, left_mean),
                ]
                lines.append("| " + " | ".join(row) + " |")

            lines.append("")
            lines.append("Project-level average (all models combined)")
            lines.append("")
            lines.append(f"| Metric | A2A | {right_label} | Δ {right_label}-A2A |")
            lines.append("| --- | --- | --- | --- |")

            def _mean_all(
                arch_data: Dict[str, Dict[str, List[float]]], arch: str
            ) -> float:
                combined: List[float] = []
                for vals in arch_data.get(arch, {}).values():
                    combined.extend(vals)
                return float(np.mean(combined)) if combined else None

            left_all = _mean_all(left_arch_data, "A2A")
            right_combined: List[float] = []
            for arch_vals in right_arch_data.values():
                for vals in arch_vals.values():
                    right_combined.extend(vals)
            right_all = float(np.mean(right_combined)) if right_combined else None
            lines.append(
                "| "
                + " | ".join(
                    [
                        "Avg tokens (all models)",
                        "-" if left_all is None else _fmt_number(left_all),
                        "-" if right_all is None else _fmt_number(right_all),
                        _fmt_delta(right_all, left_all),
                    ]
                )
                + " |"
            )

        task_set = sorted({t for t, _, _ in projects})
        for task in task_set:
            task_arch_data = arch_model_data.get(task, {})
            if not task_arch_data:
                continue
            if "A2A" not in task_arch_data or "A2A_mix" not in task_arch_data:
                continue

            left_arch_data = {"A2A": task_arch_data.get("A2A", {})}
            right_arch_data = {"A2A_mix": task_arch_data.get("A2A_mix", {})}

            lines.append("")
            lines.append(f"## {task}")

            render_pair(task, left_arch_data, right_arch_data)

        out_path = out_dir / filename
        out_dir.mkdir(parents=True, exist_ok=True)
        out_path.write_text("\n".join(lines), encoding="utf-8")
        print(f"saved markdown: {out_path}")

    write_report(
        "Pure CrewAI → MCP",
        ["Unknown", "MCP"],
        "architecture_token_deltas_crewai_to_mcp.md",
    )
    write_report("MCP → A2A", ["MCP", "A2A"], "architecture_token_deltas_mcp_to_a2a.md")
    write_a2a_to_h_a2a_report("architecture_token_deltas_a2a_to_h-a2a.md")


def generate_project_model_distribution_md(
    projects: List[Tuple[str, str, str]],
    project_data: Dict[str, Dict[str, Dict[str, List[float]]]],
    out_dir: Path,
) -> None:
    lines: List[str] = []
    lines.append("# Model token distribution per project")
    lines.append("")
    lines.append(
        "Per-project, per-model total token usage with breakdown by execution outcome. "
        "Only the mean total tokens are reported. Baseline vs maximum is computed from the overall mean aggregated across all three statuses."
    )

    def _base_task_name(task: str) -> str:
        suffixes = ("-H_A2A", "-H-A2A", "-MCP", "-A2A")
        for suffix in suffixes:
            if task.endswith(suffix):
                return task[: -len(suffix)]
        return task

    # Collect overall and series data
    overall_raw: Dict[str, List[float]] = defaultdict(list)
    series_raw: Dict[str, Dict[str, List[float]]] = defaultdict(
        lambda: defaultdict(list)
    )

    for task, arch, name in projects:
        pdata = project_data.get(name, {})
        base_task = _base_task_name(task)
        for model in MODEL_ORDER:
            for status in STATUS_ORDER:
                vals = pdata.get(status, {}).get(model, [])
                overall_raw[model].extend(vals)
                series_raw[base_task][model].extend(vals)

    def append_dist_section(
        lines: List[str],
        title: str,
        raw_data: Dict[str, List[float]],
        level_label: str = "###",
    ) -> Dict[str, float]:
        lines.append("")
        lines.append(f"{level_label} {title}")
        lines.append("| Model | n | Mean |")
        lines.append("| --- | --- | --- |")
        means: Dict[str, float] = {}
        for model in MODEL_ORDER:
            vals = raw_data.get(model, [])
            stats = summarize_values(vals)
            if stats:
                means[model] = stats["mean"]
            mean_str = _fmt_mean(stats)
            lines.append(
                f"| {MODEL_LABELS.get(model, model)} | {stats.get('count', 0)} | {mean_str} |"
            )

        lines.append("")
        lines.append(f"{level_label} Baseline vs maximum ({title})")
        if means:
            min_model = min(means.items(), key=lambda kv: kv[1])
            max_model = max(means.items(), key=lambda kv: kv[1])
            diff = max_model[1] - min_model[1]
            ratio = (
                "n/a" if min_model[1] == 0 else f"{(diff / min_model[1]) * 100:.1f}%"
            )
            lines.append(
                f"- Baseline (lowest mean): {MODEL_LABELS.get(min_model[0], min_model[0])} "
                f"= {_fmt_number(min_model[1])} tokens"
            )
            lines.append(
                f"- Maximum (highest mean): {MODEL_LABELS.get(max_model[0], max_model[0])} "
                f"= {_fmt_number(max_model[1])} tokens"
            )
            lines.append(f"- Delta: {_fmt_number(diff)} ({ratio})")
        else:
            lines.append("- No data to compare.")
        return means

    # 1. Global Overall Summary
    lines.append("")
    lines.append("## Global Summary (All Projects Combined)")
    append_dist_section(lines, "Overall per-model total tokens", overall_raw)

    # 2. Series Aggregates
    lines.append("")
    lines.append("## Series Aggregates (Aggregated by Base Task)")
    for base_task in sorted(series_raw.keys()):
        lines.append("")
        lines.append(f"### Series: {base_task}")
        append_dist_section(
            lines, f"Aggregated tokens for {base_task}", series_raw[base_task], "####"
        )

    # 3. Individual Projects
    lines.append("")
    lines.append("## Individual Project Details")
    for task, arch, name in projects:
        pdata = project_data.get(name, {})
        lines.append("")
        lines.append(f"### {name}")
        if not pdata:
            lines.append("")
            lines.append("> No token data found.")
            continue

        project_raw: Dict[str, List[float]] = {}
        for model in MODEL_ORDER:
            combined: List[float] = []
            for status in STATUS_ORDER:
                combined.extend(pdata.get(status, {}).get(model, []))
            project_raw[model] = combined

        append_dist_section(lines, "Per-model total tokens", project_raw, "####")

    out_path = out_dir / "project_model_distribution.md"
    out_dir.mkdir(parents=True, exist_ok=True)
    out_path.write_text("\n".join(lines), encoding="utf-8")
    print(f"saved markdown: {out_path}")


def plot_violin_for_project(
    project_name: str,
    project_data: Dict[str, Dict[str, List[float]]],
    out_dir: Path,
    global_max: float,
) -> None:
    if not HAS_MATPLOTLIB:
        print("matplotlib not available; skip violin plots")
        return

    any_values = False
    for status in STATUS_ORDER:
        by_model = project_data.get(status, {})
        for m in MODEL_ORDER:
            vals = by_model.get(m)
            if vals:
                any_values = True

    if not any_values or global_max <= 0.0:
        print(f"no total_tokens for project {project_name}, skip")
        return

    fig, axes = plt.subplots(1, len(STATUS_ORDER), figsize=(10, 6), sharey=True)
    if len(STATUS_ORDER) == 1:
        axes = [axes]

    y_max = global_max * 1.02
    target_ticks = 6
    while True:
        step = _nice_step(y_max, target_ticks=target_ticks)
        y_max_rounded = float(np.ceil(y_max / step) * step)
        if y_max_rounded <= y_max * 1.08 or target_ticks >= 12:
            break
        target_ticks += 2

    for idx_status, status in enumerate(STATUS_ORDER):
        ax = axes[idx_status]
        by_model = project_data.get(status, {})

        for i, model in enumerate(MODEL_ORDER, start=1):
            vals = by_model.get(model)
            if not vals:
                continue
            v_max = float(max(vals))
            parts = ax.violinplot(
                vals,
                positions=[i],
                widths=0.8,
                showmeans=False,
                showextrema=False,
                showmedians=False,
            )
            for pc in parts["bodies"]:
                pc.set_facecolor(MODEL_COLORS.get(model, "black"))
                pc.set_edgecolor("black")
                pc.set_alpha(0.7)

            median_val = float(np.median(vals))
            ax.hlines(
                median_val,
                i - 0.3,
                i + 0.3,
                colors="black",
                linewidth=1.0,
            )

            if v_max > y_max_rounded:
                y_pos = y_max_rounded * 0.985
                ax.plot([i], [y_pos], marker="^", color="black", markersize=4)
                ax.text(
                    i,
                    y_pos,
                    f">{_token_formatter(v_max, 0)}",
                    ha="center",
                    va="top",
                    fontsize=10,
                )

        ax.set_title(STATUS_TITLES.get(status, status), fontsize=24, fontweight="bold")
        ax.set_xticks(range(1, len(MODEL_ORDER) + 1))
        ax.set_xticklabels([])
        ax.set_xlim(0.5, len(MODEL_ORDER) + 0.5)
        ax.set_ylim(0, y_max_rounded)
        yticks = np.arange(0, y_max_rounded + step * 0.5, step)
        ax.set_yticks(yticks)
        ax.yaxis.set_major_formatter(FuncFormatter(_token_formatter))
        ax.grid(axis="y", linestyle="-", linewidth=0.5, alpha=0.3)
        ax.tick_params(axis="y", labelsize=22)
        ax.tick_params(axis="x", labelsize=16)
        for lbl in ax.get_yticklabels():
            lbl.set_fontweight("bold")

        if idx_status == 0:
            ax.set_ylabel("")

    fig.subplots_adjust(left=0.07, right=0.98, bottom=0.10, top=0.98, wspace=0.03)

    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / f"{project_name}_total_tokens_violin.pdf"
    fig.savefig(out_path, format="pdf", dpi=300, bbox_inches="tight", pad_inches=0.02)
    plt.close(fig)
    print(f"saved violin figure: {out_path}")


def main() -> None:
    part1_dir = Path(__file__).resolve().parent
    details_csv = part1_dir / "task_token_statistics-DETAILS.csv"
    if not details_csv.exists():
        details_csv = (
            part1_dir / "performance_reports" / "task_token_statistics-DETAILS.csv"
        )
    out_dir = part1_dir / "Violin"

    projects, project_map = load_projects(details_csv)

    project_data, arch_model_data, overall_status_values = load_all_token_data(
        details_csv, project_map
    )

    export_violin_input_summary(projects, project_data, out_dir)

    # Compute a shared y-axis maximum per task so that all architectures
    # of the same task use the same vertical scale in their violin plots.
    task_max_values: Dict[str, float] = {}
    task_values: Dict[str, List[float]] = defaultdict(list)
    for task, arch, name in projects:
        pdata = project_data.get(name, {})
        for status in STATUS_ORDER:
            by_model = pdata.get(status, {})
            for m in MODEL_ORDER:
                vals = by_model.get(m)
                if vals:
                    task_values[task].extend(vals)

    for task, vals in task_values.items():
        if not vals:
            continue
        arr = np.asarray(vals, dtype=float)
        abs_max = float(np.max(arr))
        if abs_max <= 0.0:
            continue
        task_max_values[task] = abs_max

    for task, arch, name in projects:
        pdata = project_data.get(name, {})
        global_max = task_max_values.get(task, 0.0)
        try:
            plot_violin_for_project(name, pdata, out_dir, global_max)
        except Exception as exc:
            print(f"error plotting project {name}: {exc}")

    # Generate markdown summaries
    generate_project_stats_md(projects, project_data, overall_status_values, out_dir)
    generate_architecture_deltas_md(projects, arch_model_data, out_dir)
    generate_project_model_distribution_md(projects, project_data, out_dir)


if __name__ == "__main__":
    main()