Skip to content

serejaris/personal-corp-os / skills/benchmark

benchmark

Verified

Use when choosing between engines or models for an agent pipeline and the answer must come from measurement on your own data, not from marketing pages: speech recognition for meeting recordings, the model that turns a transcript into notes, the critic model that checks it. Runs every variant in an isolated container, measures time, cost per hour of input (USD and RUB, with price source and date) and quality (WER/CER and course terms for ASR; code checks plus a judge of a different model for LLM steps), writes results.json/csv and one static analytics page. Triggers on "бенчмарк", "сравни движки", "сколько стоит час записи", "какую модель взять для пайплайна", "benchmark the pipeline", "compare ASR engines".

FileKindSize
SKILL.mdskill md7.3 KB
examples/adapter_example.pyscript4.9 KB
prices.example.jsonconfig3.3 KB
scripts/asr.pyscript10.8 KB
scripts/claude_step.pyscript5.8 KB
scripts/common.pyscript2.1 KB
scripts/llm.pyscript4.0 KB
scripts/report.pyscript10.4 KB
scripts/score.pyscript14.6 KB
task.example.jsonconfig1.7 KB

scripts/report.py

View on GitHub
#!/usr/bin/env python3
"""Render results.json as one static HTML page (tables with inline bars, light and dark).

    python3 report.py --results runs/<name>/results.json --out report.html [--title "..."] [--notes notes.md]

The page shows only aggregate numbers: no transcript text, so it is safe to share
when the recording itself is private. --notes adds a short markdown-ish list
(one bullet per line starting with "- ") under the tables.
"""

from __future__ import annotations

import argparse
import html
import json
from pathlib import Path

CSS = """
/* layout: one reading column, wide tables scroll inside their own box; bars are drawn to one scale per column */
:root{--bg:#f7f8f6;--fg:#1b211e;--muted:#5f6b65;--line:#dde3df;--bar:#2f6f8f;--bar2:#a65a3a;--ok:#2d7a4b;--bad:#b0403a;--chip:#e9eeeb;
--display:"PT Serif",Georgia,serif;--body:"PT Sans","Helvetica Neue",Arial,sans-serif;--mono:"PT Mono",ui-monospace,Menlo,monospace}
@media (prefers-color-scheme: dark){:root:not([data-theme="light"]){--bg:#141816;--fg:#e6ebe8;--muted:#98a39d;--line:#2c3430;--bar:#78b3d0;--bar2:#dd9270;--ok:#6fc392;--bad:#ef8079;--chip:#212824;color-scheme:dark}}
:root[data-theme="dark"]{--bg:#141816;--fg:#e6ebe8;--muted:#98a39d;--line:#2c3430;--bar:#78b3d0;--bar2:#dd9270;--ok:#6fc392;--bad:#ef8079;--chip:#212824;color-scheme:dark}
*{box-sizing:border-box}body{margin:0;background:var(--bg);color:var(--fg);font:15px/1.55 var(--body)}
main{max-width:1180px;margin:0 auto;padding-block:32px 64px;padding-inline:16px}
h1{font:700 28px/1.2 var(--display);margin:0 0 6px;text-wrap:balance}h2{font:700 19px/1.3 var(--display);margin:36px 0 6px;text-wrap:balance}
.sub{color:var(--muted);margin:0 0 16px;max-width:75ch}.wrap{overflow-x:auto;border-top:1px solid var(--line)}
table{border-collapse:collapse;width:100%;font-variant-numeric:tabular-nums}th,td{padding:7px 10px;border-bottom:1px solid var(--line);text-align:right;vertical-align:top;white-space:nowrap}
td{font-family:var(--mono);font-size:13px}td.l{font-family:var(--body);font-size:14px}
th{font-weight:700;color:var(--muted);font-size:11px;text-transform:uppercase;letter-spacing:.05em}td.l,th.l{text-align:left;white-space:normal;min-width:160px}
.bar{display:inline-block;height:8px;background:var(--bar);border-radius:1px;margin-left:6px;vertical-align:middle}.bar.b2{background:var(--bar2)}
.ok{color:var(--ok)}.bad{color:var(--bad)}.muted{color:var(--muted)}.chip{background:var(--chip);border-radius:3px;padding:1px 6px;font-size:12px;font-family:var(--body)}
ul{padding-left:20px;max-width:80ch}li{margin:6px 0}code{font-family:var(--mono);background:var(--chip);padding:0 4px;border-radius:3px}
"""


def esc(x) -> str:
    return html.escape("" if x is None else str(x))


def num(x, fmt="{:.2f}") -> str:
    return "—" if x is None else fmt.format(x)


def grp(n: int) -> str:
    return f"{n:,}".replace(",", "\u202f")


def bar(value, vmax, cls="") -> str:
    if value is None or not vmax:
        return ""
    return f'<span class="bar {cls}" style="width:{max(2, int(80 * value / vmax))}px"></span>'


def asr_table(rows: list[dict]) -> str:
    ok = [r for r in rows if r.get("status") == "ok"]
    wmax = max([r["wer"] for r in ok] or [0])
    cmax = max([r.get("rub_per_hour") or 0 for r in rows] or [0])
    out = ['<div class="wrap"><table><tr><th class="l">Движок</th><th>Время, мин</th><th>× реального</th>'
           '<th>₽ за час</th><th>$ за час</th><th class="l">Основа цены</th><th>WER</th><th>CER</th>'
           '<th>WER без терминов</th><th>Термины верно</th><th>Полнота терминов</th></tr>']
    for r in rows:
        if r.get("status") != "ok":
            out.append(f'<tr><td class="l">{esc(r["id"])}<br><span class="muted">{esc(r.get("error"))}</span></td>'
                       f'<td>—</td><td>—</td><td>{num(r.get("rub_per_hour"), "{:.1f}")}</td><td>{num(r.get("usd_per_hour"), "{:.3f}")}</td>'
                       f'<td class="l muted">{esc(r.get("cost_basis"))}</td><td colspan="5" class="l muted">не прогнан</td></tr>')
            continue
        out.append(
            f'<tr><td class="l">{esc(r["id"])}</td><td>{num(r["wall_s"] / 60, "{:.1f}")}</td><td>{num(r.get("rtf"), "{:.3f}")}</td>'
            f'<td>{num(r.get("rub_per_hour"), "{:.1f}")}{bar(r.get("rub_per_hour"), cmax, "b2")}</td>'
            f'<td>{num(r.get("usd_per_hour"), "{:.3f}")}</td><td class="l muted">{esc(r.get("cost_basis"))}</td>'
            f'<td>{num(r["wer"] * 100, "{:.1f}")}%{bar(r["wer"], wmax)}</td><td>{num(r["cer"] * 100, "{:.1f}")}%</td>'
            f'<td>{num(r["wer_termnorm"] * 100, "{:.1f}")}%</td>'
            f'<td>{num((r.get("term_accuracy") or 0) * 100, "{:.0f}")}% <span class="muted">({r.get("term_canon")}/{(r.get("term_canon") or 0) + (r.get("term_distortions") or 0)})</span></td>'
            f'<td>{num((r.get("term_recall") or 0) * 100, "{:.0f}")}%</td></tr>')
    out.append("</table></div>")
    return "".join(out)


def llm_table(rows: list[dict]) -> str:
    out = ['<div class="wrap"><table><tr><th class="l">Автор конспекта</th><th>Время, мин</th><th>Ходов</th>'
           '<th>Токены вход / кэш / выход</th><th>₽ за час</th><th>$ за час</th><th class="l">Проверки кодом</th>'
           '<th class="l">Судьи (critical / major / minor)</th><th>Конспект</th></tr>']
    for r in rows:
        t = r.get("tokens", {})
        first = r.get("checks_first_failed") or []
        checks = ('<span class="ok">с первого раза</span>' if r.get("checks_ok") and not first else
                  (f'<span class="ok">после {r.get("fix_rounds")} исправл.</span><br><span class="muted">сначала: {esc(", ".join(first))}</span>'
                   if r.get("checks_ok") else f'<span class="bad">не прошли</span><br><span class="muted">{esc(", ".join(r.get("checks_final_failed") or first))}</span>'))
        judges = "<br>".join(
            f'<span class="chip">{esc(j["model"] if j["model"] != "default" else "Claude")}</span> '
            f'<span class="{"ok" if j.get("verdict") == "pass" else "bad"}">{esc(j.get("verdict"))}</span> '
            f'{j.get("critical")} / {j.get("major")} / {j.get("minor")}' for j in r.get("judges", [])) or "—"
        a = r.get("artifacts", {})
        out.append(
            f'<tr><td class="l">{esc(r["author"])}<br><span class="muted">{esc(", ".join(r.get("author_models", [])))}</span></td>'
            f'<td>{num((r.get("author_wall_s") or 0) / 60, "{:.1f}")}</td><td>{esc(r.get("author_turns"))}</td>'
            f'<td>{grp(t.get("input_tokens", 0))} / {grp(t.get("cache_read_input_tokens", 0) + t.get("cache_creation_input_tokens", 0))} / {grp(t.get("output_tokens", 0))}</td>'
            f'<td>{num(r.get("rub_per_hour"), "{:.1f}")}</td><td>{num(r.get("usd_per_hour"), "{:.3f}")}</td>'
            f'<td class="l">{checks}</td><td class="l">{judges}</td>'
            f'<td>{esc(a.get("lesson_kb"))} КБ, {esc(a.get("sections"))} разд., {esc(a.get("chapters"))} глав</td></tr>')
    out.append("</table></div>")
    return "".join(out).replace(",", " ").replace("  ", ", ")


def pairwise_table(pw: dict) -> str:
    if not pw:
        return ""
    ids = list(pw)
    out = ['<h2>Движки друг против друга</h2><p class="sub">WER, если текст движка в столбце считать эталоном для движка в строке. '
           'Не зависит от эталонного транскрипта.</p><div class="wrap"><table><tr><th class="l"></th>']
    out += [f"<th>{esc(i)}</th>" for i in ids] + ["</tr>"]
    for a in ids:
        out.append(f'<tr><td class="l">{esc(a)}</td>')
        out += [f"<td>{'—' if a == b else num(pw[a].get(b, 0) * 100, '{:.1f}') + '%'}</td>" for b in ids]
        out.append("</tr>")
    out.append("</table></div>")
    return "".join(out)


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--results", required=True, type=Path)
    ap.add_argument("--out", required=True, type=Path)
    ap.add_argument("--title", default="Бенчмарк пайплайна")
    ap.add_argument("--notes", type=Path)
    ap.add_argument("--fragment", action="store_true", help="no doctype/html/body wrapper (for hosts that add their own)")
    args = ap.parse_args()
    res = json.loads(args.results.read_text(encoding="utf-8"))
    audio_min = (res.get("audio_s") or 0) / 60
    fx = res.get("fx", {})
    notes = ""
    if args.notes and args.notes.exists():
        items = [l[2:].strip() for l in args.notes.read_text(encoding="utf-8").splitlines() if l.startswith("- ")]
        notes = "<h2>Как читать</h2><ul>" + "".join(f"<li>{esc(i)}</li>" for i in items) + "</ul>"
    fonts = ('<link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=PT+Mono&family=PT+Sans:wght@400;700'
             '&family=PT+Serif:wght@700&display=swap">')
    head = f"<title>{esc(args.title)}</title>{fonts}<style>{CSS}</style>"
    body = f"""<main>
<h1>{esc(args.title)}</h1>
<p class="sub">Задача <code>{esc(res.get("task"))}</code> · запись {audio_min:.0f} мин · цены на {esc(res.get("prices_checked_at"))},
курс ЦБ {esc(fx.get("USD_RUB"))} ₽ за $ ({esc(fx.get("date"))}) · собрано {esc(res.get("generated_at"))}</p>
<h2>Распознавание: звук в текст</h2>
<p class="sub">Цена и время приведены к часу записи. WER и CER считаются против эталонного транскрипта ({esc(res.get("reference", {}).get("words"))} слов); меньше лучше.</p>
{asr_table(res.get("asr", []))}
<h2>Разбор: транскрипт в конспект</h2>
<p class="sub">Один и тот же транскрипт, один харнес (Claude Code, headless), меняется только модель. Судья всегда другой модели, чем автор.</p>
{llm_table(res.get("llm", []))}
{pairwise_table(res.get("asr_pairwise_wer", {}))}
{notes}
</main>
"""
    page = head + body if args.fragment else (
        '<!doctype html><html lang="ru"><head><meta charset="utf-8">'
        '<meta name="viewport" content="width=device-width,initial-scale=1">' + head + "</head><body>" + body + "</body></html>")
    args.out.write_text(page, encoding="utf-8")
    print(f"report: {args.out}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall