bench_t2v_camera_time/summarize.py

# SPDX-License-Identifier: Apache-2.0
"""Aggregate score_claude.py output into a per-video score file and a summary.

Replaces the old combine.py, which merged two judges (Qwen3.5 + VideoScore2).
Claude Opus 5 is now the only judge, so there is nothing to combine -- the
score is just the mean of the four rubric axes:

  score = mean(time_alignment, camera_motion, quality, smoothness) / 10
  pass  = time_alignment >= 8 and camera_motion >= 6 and quality >= 7

Thresholds are carried over unchanged from combine.py (camera_motion stays at
6, not 8, since camera-motion-following is a harder, less-established
capability than the transition/quality axes); the VideoScore2 `min(v,t,p) >= 4`
gate is gone with the model that produced it.

Usage:
    python summarize.py
    python summarize.py --out-dir outputs
Output: outputs/scores.jsonl (per-video) + outputs/summary.json.
"""
from __future__ import annotations

import argparse
import json
from pathlib import Path

import numpy as np

HERE = Path(__file__).resolve().parent

AXES = ("time_alignment", "camera_motion", "quality", "smoothness")
GROUP_KEYS = ("camera_motion_name", "time_variant", "domain")

PASS_TIME_ALIGNMENT = 8.0
PASS_CAMERA_MOTION = 6.0
PASS_QUALITY = 7.0


def passed(r: dict) -> bool:
    return (r["time_alignment"] >= PASS_TIME_ALIGNMENT
            and r["camera_motion"] >= PASS_CAMERA_MOTION
            and r["quality"] >= PASS_QUALITY)


# --- generic below this line -------------------------------------------------

def parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(description="Summarize Claude judge scores.")
    p.add_argument("--out-dir", default=str(HERE / "outputs"))
    return p.parse_args()


def load_jsonl(path: Path) -> list[dict]:
    if not path.exists():
        raise SystemExit(f"missing {path} -- run score_claude.py first")
    return [json.loads(l) for l in path.read_text().splitlines() if l.strip()]


def main() -> None:
    args = parse_args()
    out_dir = Path(args.out_dir)
    records = load_jsonl(out_dir / "claude_scores.jsonl")

    merged = []
    for r in sorted(records, key=lambda x: x["id"]):
        if "error" in r:
            merged.append({**{k: v for k, v in r.items() if k != "reason"}, "pass": False})
            continue
        score = sum(r[axis] for axis in AXES) / (10.0 * len(AXES))
        merged.append({**r, "score": round(score, 4), "pass": passed(r)})

    scores_path = out_dir / "scores.jsonl"
    scores_path.write_text("\n".join(json.dumps(r) for r in merged) + "\n")

    ok = [r for r in merged if "error" not in r]
    summary = {
        "judge": "claude-opus-5 (bedrock)",
        "num_scored": len(merged),
        "num_ok": len(ok),
        "num_errors": len(merged) - len(ok),
        "pass_rate": round(sum(r["pass"] for r in ok) / len(ok), 3) if ok else None,
        "mean_score": round(float(np.mean([r["score"] for r in ok])), 3) if ok else None,
    }
    for axis in AXES:
        summary[f"mean_{axis}"] = round(float(np.mean([r[axis] for r in ok])), 3) if ok else None
    for key in GROUP_KEYS:
        groups: dict = {}
        for r in ok:
            groups.setdefault(r[key], []).append(r["score"])
        summary[f"mean_score_by_{key}"] = {k: round(float(np.mean(v)), 3)
                                           for k, v in sorted(groups.items())}

    (out_dir / "summary.json").write_text(json.dumps(summary, indent=2))
    print(json.dumps(summary, indent=2))


if __name__ == "__main__":
    main()