step_train/scripts_train/analyze_kvlen.py

"""Cross-cell results aggregator for the trainable-KV length ablation.

Run AFTER the 10 cells (5 lengths × 2 datasets) finish. Walks every
data/eval/kvlen/<DS>/p<P>/preds.jsonl, recomputes metrics from the per-record
source of truth (so it's robust to a cell that was interrupted mid-judge), and
renders side-by-side comparison tables that no single per-cell stats.json gives:

  1. Per-dataset summary: rows = KV length p, cols = n / coverage / judge_acc /
     judge_correct_rate / em / f1
  2. Per-dataset judge_acc broken out BY query_type × KV length (which question
     types benefit from a longer cartridge?)

Pure stdlib — NO torch / NO cartridges import — so it runs on any machine
(including CPU-only), unlike eval_direct_ask.py.

Usage:
    python scripts/train/analyze_kvlen.py                       # both datasets, default lengths
    python scripts/train/analyze_kvlen.py --datasets lmes       # one dataset
    python scripts/train/analyze_kvlen.py --csv data/eval/kvlen/summary.csv
"""

import argparse
import json
import os
import sys
from collections import defaultdict

PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

DEFAULT_DATASETS = ["lmes", "metamem5k"]
DEFAULT_LENGTHS = [64, 128, 256, 512, 1024]
DS_DIR = {"lmes": "longmemeval_s", "metamem5k": "metamem_5k"}


def _r(p):
    return p if os.path.isabs(p) else os.path.join(PROJECT_ROOT, p)


# ---- metric helpers (verdict->numeric mapping kept identical to ----
# ---- eval_direct_ask.aggregate_direct / eval_metrics._judge_num) ----
def _judge_num(v):
    if isinstance(v, str):
        return {"correct": 1.0, "partial": 0.5, "wrong": 0.0}.get(v)
    if isinstance(v, bool):
        return 1.0 if v else 0.0
    return None


def _mean(xs):
    xs = [x for x in xs if x is not None]
    return sum(xs) / len(xs) if xs else None


def _read_jsonl(path):
    recs = []
    with open(path) as f:
        for line in f:
            line = line.strip()
            if line:
                try:
                    recs.append(json.loads(line))
                except Exception:
                    pass
    return recs


def analyze_cell(eval_root, ds, p, n_sampled=None):
    """Recompute one cell's metrics from preds.jsonl. Returns None if not run yet."""
    preds = os.path.join(_r(eval_root), ds, f"p{p}", "preds.jsonl")
    if not os.path.exists(preds):
        return None
    recs = _read_jsonl(preds)
    if not recs:
        return {"n": 0, "exists": True}

    jc = [1.0 if r.get("answer_judge") == "correct" else 0.0
          for r in recs if r.get("answer_judge") is not None]
    users_answered = {r.get("user_id") for r in recs}
    cell = {
        "exists": True,
        "n": len(recs),
        "n_judged": len(jc),
        "n_users": len(users_answered),
        "coverage": (len(users_answered) / n_sampled) if n_sampled else None,
        "judge_acc": _mean([_judge_num(r.get("answer_judge")) for r in recs]),
        "judge_correct_rate": (sum(jc) / len(jc)) if jc else None,
        "em": _mean([r.get("answer_em") for r in recs]),
        "f1": _mean([r.get("answer_f1") for r in recs]),
    }
    by = defaultdict(list)
    for r in recs:
        by[r.get("query_type")].append(r)
    cell["by_qtype"] = {
        str(qt): _mean([_judge_num(r.get("answer_judge")) for r in rs])
        for qt, rs in by.items()
    }
    cell["by_qtype_n"] = {str(qt): len(rs) for qt, rs in by.items()}
    return cell


def _load_sample_count(ds, seed):
    """How many users were sampled (denominator for coverage). None if list missing."""
    path = os.path.join(
        _r("data/processed"), DS_DIR[ds], "splits", f"kvlen_sample200_seed{seed}.json"
    )
    if os.path.exists(path):
        try:
            return len(json.load(open(path)))
        except Exception:
            return None
    return None


def _fmt(x, pct=False):
    if x is None:
        return "  — "
    if pct:
        return f"{100*x:5.1f}"
    return f"{x:5.3f}" if isinstance(x, float) else str(x)


def print_summary(ds, lengths, cells, n_sampled):
    print(f"\n{'='*78}")
    print(f"  Dataset: {ds}   (sampled users: {n_sampled if n_sampled else '?'})")
    print(f"{'='*78}")
    hdr = f"{'p (KV len)':>10} | {'n':>5} {'users':>6} {'cov%':>6} | " \
          f"{'judge_acc':>9} {'correct%':>9} | {'EM':>6} {'F1':>6}"
    print(hdr)
    print("-" * len(hdr))
    for p in lengths:
        c = cells.get(p)
        if c is None:
            print(f"{('p'+str(p)):>10} | {'(not run yet)':>30}")
            continue
        print(
            f"{('p'+str(p)):>10} | {c['n']:>5} {c.get('n_users','-'):>6} "
            f"{_fmt(c.get('coverage'), pct=True):>6} | "
            f"{_fmt(c.get('judge_acc')):>9} {_fmt(c.get('judge_correct_rate'), pct=True):>9} | "
            f"{_fmt(c.get('em')):>6} {_fmt(c.get('f1')):>6}"
        )


def print_by_qtype(ds, lengths, cells):
    """judge_acc broken out by query_type (rows) × KV length (cols)."""
    qtypes = sorted({qt for c in cells.values() if c for qt in c.get("by_qtype", {})})
    if not qtypes:
        return
    print(f"\n  [{ds}] judge_acc by query_type × KV length")
    cols = " ".join(f"{('p'+str(p)):>7}" for p in lengths)
    print(f"  {'query_type':>26} | {cols}")
    print(f"  {'-'*26}-+-{'-'*len(cols)}")
    for qt in qtypes:
        row = []
        for p in lengths:
            c = cells.get(p)
            v = c.get("by_qtype", {}).get(qt) if c else None
            row.append(f"{_fmt(v):>7}")
        print(f"  {qt:>26} | {' '.join(row)}")


def main():
    ap = argparse.ArgumentParser(description="Cross-cell KV-length ablation analyzer")
    ap.add_argument("--datasets", nargs="+", default=DEFAULT_DATASETS,
                    choices=DEFAULT_DATASETS)
    ap.add_argument("--lengths", nargs="+", type=int, default=DEFAULT_LENGTHS)
    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--eval-root", default="data/eval/kvlen")
    ap.add_argument("--csv", default=None, help="Optional: write a flat CSV for plotting")
    ap.add_argument("--no-qtype", action="store_true", help="Skip the by-query_type table")
    args = ap.parse_args()

    all_results = {}  # ds -> {p -> cell}
    for ds in args.datasets:
        n_sampled = _load_sample_count(ds, args.seed)
        cells = {}
        for p in args.lengths:
            cells[p] = analyze_cell(args.eval_root, ds, p, n_sampled=n_sampled)
        all_results[ds] = (cells, n_sampled)
        print_summary(ds, args.lengths, cells, n_sampled)
        if not args.no_qtype:
            print_by_qtype(ds, args.lengths, cells)

    # ---- optional CSV (one row per cell; flat, plot-friendly) ----
    if args.csv:
        csv_path = _r(args.csv)
        os.makedirs(os.path.dirname(csv_path), exist_ok=True)
        with open(csv_path, "w", encoding="utf-8") as f:
            f.write("dataset,kv_len,n,n_users,n_sampled,coverage,judge_acc,"
                    "judge_correct_rate,em,f1\n")
            for ds in args.datasets:
                cells, n_sampled = all_results[ds]
                for p in args.lengths:
                    c = cells.get(p)
                    if not c:
                        continue
                    f.write(",".join(str(x) for x in [
                        ds, p, c.get("n", 0), c.get("n_users", ""),
                        n_sampled if n_sampled else "", c.get("coverage", ""),
                        c.get("judge_acc", ""), c.get("judge_correct_rate", ""),
                        c.get("em", ""), c.get("f1", ""),
                    ]) + "\n")
        print(f"\nCSV written -> {csv_path}")

    # ---- machine-readable JSON dump alongside CSV semantics ----
    missing = [(ds, p) for ds in args.datasets
               for p in args.lengths if all_results[ds][0].get(p) is None]
    if missing:
        print(f"\n[note] {len(missing)} cell(s) not run yet: "
              + ", ".join(f"{ds}/p{p}" for ds, p in missing))


if __name__ == "__main__":
    main()