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/asr.py

View on GitHub
#!/usr/bin/env python3
"""Run the ASR variants of a benchmark task, one after another.

    python3 asr.py --task task.json --out runs/<name> [--only id1,id2] [--force]

For every variant writes runs/<name>/asr/<id>/:
    transcript.txt   plain text the engine returned
    raw.json         full engine response (words, timings, speakers, usage)
    meta.json        engine, model, wall_s, cpu_s, audio_s, status, error, usage

Engines (field "engine" of a variant):
    elevenlabs       ElevenLabs Speech to Text (Scribe). "keyterms": true sends the
                     task terms list as the keyterms field. Credits used are read from
                     /v1/user/subscription before and after the call.
    openai_compat    POST <base_url>/audio/transcriptions (OpenAI-compatible server:
                     self-hosted GigaAM, speaches, vLLM, a cloud endpoint).
    faster_whisper   Whisper on the local CPU through faster-whisper (CTranslate2).
    openrouter_audio chat completions with input_audio chunks (models with audio input
                     on OpenRouter or any OpenAI-compatible chat endpoint).

Keys are read from env only (names in the variant: "key_env"); values are never logged.
Variants run sequentially so wall time and CPU time are not shared between engines.
"""

from __future__ import annotations

import argparse
import base64
import json
import re
import subprocess
import sys
import tempfile
import traceback
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent))
from common import Meter, audio_seconds, load_task, read_json, secret, task_path, utc_now, write_json  # noqa: E402

KEYLIKE = re.compile(r"[A-Za-z0-9_\-]{24,}")


def redact(text: str) -> str:
    """Error bodies of some providers echo the key back: never keep long tokens."""
    return KEYLIKE.sub("***", text)[:400]


# --- terms -------------------------------------------------------------------

def load_keyterms(task: dict) -> list[str]:
    path = task_path(task, task.get("terms"))
    if not path:
        return []
    data = read_json(path)
    out, seen = [], set()
    for t in data.get("terms", []):
        if t.get("keyterm") is False:
            continue
        for k in t.get("keyterms") or [t.get("canon")]:
            k = " ".join(str(k or "").split())
            # Scribe limits: <50 chars, <=5 words, no <>{}[]\ ; up to 1000 terms
            if not k or k.lower() in seen or len(k) >= 50 or len(k.split()) > 5 or set(k) & set("<>{}[]\\"):
                continue
            seen.add(k.lower())
            out.append(k)
    return out[:1000]


# --- engines -----------------------------------------------------------------

def run_elevenlabs(v: dict, audio: Path, task: dict) -> tuple[str, dict, dict]:
    import httpx

    key = secret(v.get("key_env", "ELEVENLABS_API_KEY"))
    base = v.get("base_url", "https://api.elevenlabs.io/v1")
    headers = {"xi-api-key": key}

    def credits() -> int | None:
        try:
            r = httpx.get(f"{base}/user/subscription", headers=headers, timeout=30)
            return r.json().get("character_count") if r.status_code == 200 else None
        except httpx.HTTPError:
            return None

    data: dict = {
        "model_id": v.get("model", "scribe_v2"),
        "language_code": task.get("language", "ru"),
        "timestamps_granularity": "word",
        "diarize": "true",
    }
    keyterms = load_keyterms(task) if v.get("keyterms") else []
    if keyterms:
        data["keyterms"] = keyterms
    before = credits()
    with audio.open("rb") as fh, httpx.Client(timeout=httpx.Timeout(3600, connect=30)) as client:
        r = client.post(f"{base}/speech-to-text", headers=headers, data=data, files={"file": (audio.name, fh)})
    if r.status_code != 200:
        raise RuntimeError(f"HTTP {r.status_code}: {redact(r.text)}")
    payload = r.json()
    after = credits()
    usage = {"keyterms": len(keyterms), "credits_before": before, "credits_after": after,
             "credits_used": (after - before) if before is not None and after is not None else None}
    return payload.get("text", ""), payload, usage


def run_openai_compat(v: dict, audio: Path, task: dict) -> tuple[str, dict, dict]:
    import httpx

    headers = {}
    if v.get("key_env"):
        headers["Authorization"] = f"Bearer {secret(v['key_env'])}"
    data = {"model": v["model"]}
    if v.get("send_language", True) and task.get("language"):
        data["language"] = task["language"]
    with audio.open("rb") as fh, httpx.Client(timeout=httpx.Timeout(float(v.get("timeout", 7200)), connect=30)) as client:
        r = client.post(v["base_url"].rstrip("/") + "/audio/transcriptions", headers=headers,
                        data=data, files={"file": (audio.name, fh)})
    if r.status_code != 200:
        raise RuntimeError(f"HTTP {r.status_code}: {redact(r.text)}")
    payload = r.json()
    return payload.get("text", ""), payload, {}


def run_faster_whisper(v: dict, audio: Path, task: dict) -> tuple[str, dict, dict]:
    from faster_whisper import WhisperModel

    model = WhisperModel(v.get("model", "large-v3-turbo"), device="cpu",
                         compute_type=v.get("compute_type", "int8"), cpu_threads=int(v.get("threads", 8)))
    prompt = None
    if v.get("terms_prompt"):
        prompt = ", ".join(load_keyterms(task))[:800]
    # Decode with ffmpeg ourselves: faster-whisper's PyAV path breaks on some av releases.
    import numpy as np

    pcm = subprocess.run(["ffmpeg", "-hide_banner", "-loglevel", "error", "-i", str(audio), "-ac", "1", "-ar", "16000",
                          "-f", "s16le", "-"], capture_output=True, check=True).stdout
    samples = np.frombuffer(pcm, np.int16).astype(np.float32) / 32768.0
    segments, info = model.transcribe(samples, language=task.get("language", "ru"),
                                      beam_size=int(v.get("beam_size", 5)), vad_filter=bool(v.get("vad", True)),
                                      initial_prompt=prompt)
    segs = [{"start": round(s.start, 2), "end": round(s.end, 2), "text": s.text.strip()} for s in segments]
    text = " ".join(s["text"] for s in segs)
    return text, {"language": info.language, "duration": info.duration, "segments": segs}, {"terms_prompt": bool(prompt)}


def run_openrouter_audio(v: dict, audio: Path, task: dict) -> tuple[str, dict, dict]:
    import httpx

    key = secret(v.get("key_env", "OPENROUTER_API_KEY"))
    base = v.get("base_url", "https://openrouter.ai/api/v1").rstrip("/")
    minutes = float(v.get("chunk_minutes", 10))
    prompt = v.get("prompt") or (
        "Transcribe this audio verbatim in its original language. Output only the transcript text, "
        "no comments, no timestamps, no speaker labels.")
    if v.get("terms_prompt"):
        prompt += " Spelling of names and terms: " + ", ".join(load_keyterms(task))
    texts, calls = [], []
    with tempfile.TemporaryDirectory() as td:
        pattern = Path(td) / "chunk-%03d.mp3"
        subprocess.run(["ffmpeg", "-hide_banner", "-loglevel", "error", "-y", "-i", str(audio), "-ac", "1",
                        "-ar", "16000", "-b:a", "48k", "-f", "segment", "-segment_time", str(int(minutes * 60)),
                        str(pattern)], check=True)
        for chunk in sorted(Path(td).glob("chunk-*.mp3")):
            body = {
                "model": v["model"],
                "messages": [{"role": "user", "content": [
                    {"type": "text", "text": prompt},
                    {"type": "input_audio", "input_audio": {"data": base64.b64encode(chunk.read_bytes()).decode(),
                                                            "format": "mp3"}},
                ]}],
                "temperature": 0,
                "usage": {"include": True},
            }
            r = httpx.post(f"{base}/chat/completions", headers={"Authorization": f"Bearer {key}"}, json=body,
                           timeout=httpx.Timeout(1800, connect=30))
            if r.status_code != 200:
                raise RuntimeError(f"{chunk.name}: HTTP {r.status_code}: {redact(r.text)}")
            p = r.json()
            texts.append((p["choices"][0]["message"].get("content") or "").strip())
            calls.append({"chunk": chunk.name, "usage": p.get("usage", {}), "model": p.get("model")})
    usage = {"calls": len(calls),
             "prompt_tokens": sum(c["usage"].get("prompt_tokens", 0) for c in calls),
             "completion_tokens": sum(c["usage"].get("completion_tokens", 0) for c in calls),
             "cost_usd_reported": round(sum(float(c["usage"].get("cost") or 0) for c in calls), 6)}
    return "\n".join(texts), {"calls": calls}, usage


ENGINES = {
    "elevenlabs": run_elevenlabs,
    "openai_compat": run_openai_compat,
    "faster_whisper": run_faster_whisper,
    "openrouter_audio": run_openrouter_audio,
}


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--task", required=True, type=Path)
    ap.add_argument("--out", required=True, type=Path)
    ap.add_argument("--only", help="comma-separated variant ids")
    ap.add_argument("--force", action="store_true", help="rerun variants that already have status ok")
    args = ap.parse_args()

    task = load_task(args.task)
    audio = task_path(task, task["audio"])
    audio_s = audio_seconds(audio)
    only = set(args.only.split(",")) if args.only else None
    for v in task.get("asr", []):
        if only and v["id"] not in only:
            continue
        if v.get("skip"):
            continue
        out = args.out / "asr" / v["id"]
        meta_path = out / "meta.json"
        if meta_path.exists() and read_json(meta_path).get("status") == "ok" and not args.force:
            print(f"{v['id']}: already ok, skip")
            continue
        out.mkdir(parents=True, exist_ok=True)
        meta = {"id": v["id"], "engine": v["engine"], "model": v.get("model"), "audio_s": round(audio_s, 1),
                "started_at": utc_now(), "variant": {k: val for k, val in v.items() if k != "prompt"}}
        print(f"{v['id']}: start", flush=True)
        try:
            with Meter() as m:
                text, raw, usage = ENGINES[v["engine"]](v, audio, task)
            (out / "transcript.txt").write_text(text.strip() + "\n", encoding="utf-8")
            write_json(out / "raw.json", raw)
            meta.update(status="ok", wall_s=m.wall_s, cpu_s=m.cpu_s, usage=usage, chars=len(text))
        except Exception as exc:  # one failed engine must not stop the others
            meta.update(status="error", error=redact(f"{type(exc).__name__}: {exc}"),
                        trace=redact(traceback.format_exc().splitlines()[-1]))
        meta["finished_at"] = utc_now()
        write_json(meta_path, meta)
        print(f"{v['id']}: {meta['status']} {meta.get('wall_s', '')}s {meta.get('error', '')}", flush=True)
    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