scripts/run_benchmark.py

#!/usr/bin/env python3
"""Run the Entity Transcription Benchmark against any STT endpoint.

    export DEEPGRAM_API_KEY=...
    python run_benchmark.py --provider deepgram --provider-opt model=nova-3 \
        --out runs/deepgram_nova3.jsonl

    export MY_KEY=...
    python run_benchmark.py --provider http --endpoint https://api.example.com/v1/stt \
        --api-key-env MY_KEY --response-path result.transcript --out runs/mine.jsonl

Output is a resumable JSONL of hypotheses plus a manifest recording exactly how
the run was configured. Scoring is a separate step: score_benchmark.py.
"""

from __future__ import annotations

import argparse
import json
import os
import sys
import time
from concurrent.futures import ThreadPoolExecutor, as_completed
from datetime import datetime, timezone

import common
import providers


def parse_args(argv=None):
    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    p.add_argument("--provider", default="http", choices=sorted(providers.PROVIDERS),
                   help="'http' (default) works with any endpoint that takes audio and "
                        "returns JSON. The named adapters are conveniences, not requirements.")
    p.add_argument("--system-name", default=None,
                   help="label for the system under test, recorded in the manifest and "
                        "summary (e.g. 'acme-stt-v2'). Defaults to the provider name.")
    p.add_argument("--endpoint", help="for --provider http, or to override a vendor base URL")
    p.add_argument("--api-key-env", default=None,
                   help="env var holding the API key (default: <PROVIDER>_API_KEY)")
    p.add_argument("--provider-opt", action="append", default=[], metavar="K=V",
                   help="repeatable; passed to the adapter (e.g. model=nova-3 smart_format=true)")
    p.add_argument("--response-path", help="shorthand for --provider-opt response_path=...")

    p.add_argument("--dataset", default=common.DEFAULT_DATASET)
    p.add_argument("--split", default=common.DEFAULT_SPLIT)
    p.add_argument("--revision", default=None, help="pin a dataset commit SHA -- recommended")
    p.add_argument("--local-path", default=None, help="load_from_disk path instead of the Hub")
    p.add_argument("--tier", default="all",
                   help="skip clips with no span of this tier. Default 'all' -- transcribe "
                        "every clip, since WER is reported over the full split and the tier "
                        "filter is applied per span at scoring time.")
    p.add_argument("--limit", type=int, default=0, help="first N clips; smoke tests only")

    p.add_argument("--out", required=True, help="hypotheses JSONL")
    p.add_argument("--workers", type=int, default=4)
    p.add_argument("--timeout", type=int, default=providers.DEFAULT_TIMEOUT)
    p.add_argument("--retries", type=int, default=3)
    p.add_argument("--overwrite", action="store_true", help="ignore existing output and start over")

    for field in common.CANDIDATES:
        p.add_argument(f"--{field.replace('_', '-')}-column", dest=f"{field}_column", default=None)
    return p.parse_args(argv)


def build_cfg(args) -> dict:
    opts: dict = {}
    for item in args.provider_opt:
        key, _, value = item.partition("=")
        opts[key.strip()] = value.strip()
    if args.endpoint:
        opts["endpoint"] = args.endpoint
    if args.response_path:
        opts["response_path"] = args.response_path
    key_env = args.api_key_env or f"{args.provider.upper()}_API_KEY"
    return {
        "provider": args.provider,
        "system_name": args.system_name or args.provider,
        "api_key_env": key_env,
        "opts": opts,
        "timeout": args.timeout,
    }


def load_done(path: str) -> dict[str, dict]:
    if not os.path.exists(path):
        return {}
    done = {}
    with open(path) as fh:
        for line in fh:
            line = line.strip()
            if not line:
                continue
            try:
                record = json.loads(line)
            except json.JSONDecodeError:
                continue
            if record.get("hypothesis") is not None and not record.get("error"):
                done[record["clip_id"]] = record
    return done


def main(argv=None) -> int:
    args = parse_args(argv)
    cfg = build_cfg(args)

    anonymous_ok = args.provider in ("mock", "http") and not args.api_key_env
    if not anonymous_ok and not os.environ.get(cfg["api_key_env"]):
        common.eprint(
            f"! {cfg['api_key_env']} is not set. Export your key, or pass --api-key-env "
            f"to name a different variable."
        )
        return 2
    if args.provider == "http" and not args.endpoint:
        common.eprint("! --provider http needs --endpoint")
        return 2

    ds = common.load_bench(args.dataset, args.split, args.revision, args.local_path)
    overrides = {f: getattr(args, f"{f}_column") for f in common.CANDIDATES}
    schema = common.resolve_schema(ds.column_names, overrides)
    common.eprint("Resolved columns:\n" + schema.pretty())

    if args.tier and args.tier.lower() != "all":
        ds = ds.filter(lambda row: common.row_has_tier(row, schema, args.tier))
    ds = common.undecode_audio(ds, schema)
    if args.limit:
        ds = ds.select(range(min(args.limit, len(ds))))
    common.eprint(f"Clips to transcribe: {len(ds)} (tier={args.tier})")

    os.makedirs(os.path.dirname(os.path.abspath(args.out)) or ".", exist_ok=True)
    if args.overwrite and os.path.exists(args.out):
        os.remove(args.out)
    done = load_done(args.out)
    if done:
        common.eprint(f"Resuming: {len(done)} clips already transcribed")

    manifest = {
        "harness_version": "1.0.0",
        "started_utc": datetime.now(timezone.utc).isoformat(timespec="seconds"),
        "dataset": args.dataset if not args.local_path else f"local:{args.local_path}",
        "split": args.split,
        "revision": args.revision,
        "tier": args.tier,
        "system_name": cfg["system_name"],
        "provider": args.provider,
        "provider_opts": cfg["opts"],
        "api_key_env": cfg["api_key_env"],
        "clips": len(ds),
    }
    manifest_path = os.path.splitext(args.out)[0] + ".manifest.json"
    with open(manifest_path, "w") as fh:
        json.dump(manifest, fh, indent=2)

    pending = []
    for index, row in enumerate(ds):
        clip = common.row_id(row, schema, index)
        if clip in done:
            continue
        pending.append((index, clip, row))

    if not pending:
        common.eprint("Nothing to do -- all clips already present.")
        return 0

    started = time.time()
    errors = 0
    sink = open(args.out, "a", buffering=1)

    def work(item):
        index, clip, row = item
        try:
            audio, sample_rate = common.audio_to_wav_bytes(row[schema.audio])
            call_cfg = dict(cfg)
            call_cfg["clip_id"] = clip
            hypothesis = providers.transcribe(audio, call_cfg, retries=args.retries)
            return {
                "clip_id": clip,
                "index": index,
                "hypothesis": hypothesis,
                "sampling_rate": sample_rate,
            }
        except Exception as exc:  # noqa: BLE001 -- one bad clip must not kill the run
            return {"clip_id": clip, "index": index, "hypothesis": None, "error": str(exc)}

    try:
        with ThreadPoolExecutor(max_workers=args.workers) as pool:
            futures = {pool.submit(work, item): item for item in pending}
            for completed, future in enumerate(as_completed(futures), 1):
                record = future.result()
                if record.get("error"):
                    errors += 1
                    if errors <= 5:
                        common.eprint(f"  ! {record['clip_id']}: {record['error'][:160]}")
                    if errors == 6:
                        common.eprint("  ! (further errors suppressed; see the JSONL)")
                sink.write(json.dumps(record, ensure_ascii=False) + "\n")
                if completed % 25 == 0 or completed == len(pending):
                    rate = completed / max(time.time() - started, 1e-6)
                    common.eprint(
                        f"  {completed}/{len(pending)}  {rate:.1f} clips/s  errors={errors}"
                    )
    except KeyboardInterrupt:
        common.eprint("\nInterrupted -- rerun the same command to resume.")
        return 130
    finally:
        sink.close()

    if errors:
        common.eprint(
            f"\nDone with {errors} failed clips. Rerun the same command to retry only those."
        )
    common.eprint(f"\nHypotheses: {args.out}\nManifest:   {manifest_path}")
    common.eprint(f"Next: python score_benchmark.py --hyps {args.out}")
    return 1 if errors else 0


if __name__ == "__main__":
    sys.exit(main())