Blame for code/mteb_lite_select_v2.py

#!/usr/bin/env python
"""MTEB-lite v2: IRT / anchor-point item selection on a DIVERSE reference set, validated by family-level LOMO.

Why v2: with 33 of 42 reference models being near-identical Qwen merges, v1's "discriminative" items (binned MI,
mRMR) chased noise and lost to random K under leave-one-model-out. v2:
  1. Reference models are grouped into families (by merge recipe; every public / base model is its own family)
     and thinned to a diverse set: <= --per-family models per family (best / median / worst by full score).
  2. Per task, item outcomes are modelled against the known model ability theta (= full-task score standardised
     over the training models):
       continuous items  linear-Gaussian factor model  y_im = a_i theta_m + b_i + e,  e ~ N(0, s_i^2)
       binary items      2PL  P(y_im = 1) = sigmoid(a_i (theta_m - d_i))  (L2-regularised, vectorised Newton)
     Selection (per-item tasks: retrieval, reranking, classification, bitext, summarization):
       anchor   tinyBenchmarks anchor points: k-means (K clusters) of the items' outcome vectors over the training
                models; anchor = item nearest each centroid, weight = the cluster's share of the pool (subset-aware)
       fisher   greedy max of sum_m log(eps + I_S(theta_m)) (information spread over the reference abilities)
     Estimators of the full score (all finally mapped by a linear calibration fitted on the training models):
       mean     (anchor-weighted) subset mean
       pirt     p-IRT: MAP theta from the selected items, pool score = observed items + model-predicted rest
       gp       0.5 * (mean + pirt)
     Set-metric tasks (STS Spearman, clustering V-measure, pair-classification AP): no per-item IRT applies (the
     metric is not a sum over items) -> stratified random K (gold-score bins / labels / classes, per subset) +
     calibration.
  3. Validation: family-level LOMO (hold out a whole family; the honest generalisation test) and distinct-model
     LOMO (hold out one model of the diverse set), K=--K; baselines: uniform random K (metrics averaged over
     --n-random single draws - NOT averaged predictions, which v1 did and which made "random" an ensemble of
     5K items) and the whole pool.

  python scripts/mteb_lite_select_v2.py --K 100 [--suite eng|ara] [--jobs 128] [--full-csv data/mteb_lite/public_full.csv]
Writes <out>/selection/<task>.json, <out>/<manifest> and <report>.{md,json}
(defaults: data/mteb_lite_v2, outputs/mteb_lite/selection_report_v2; ara: data/mteb_lite_ar_v2, ...).
"""
from __future__ import annotations

import argparse
import json
import sys
import time
from collections import defaultdict
from pathlib import Path

import numpy as np

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))
import mteb_lite_lib as L  # noqa: E402
import mteb_lite_select as S1  # noqa: E402  (full scores, metrics)

REPO = L.REPO
PER_ITEM = {"Retrieval", "Reranking", "Classification", "BitextMining", "Summarization"}
METHODS = ("anchor", "fisher")
ESTIMATORS = ("mean", "pirt", "gp")


# ----------------------------------------------------------------------------- reference models
def family(name: str) -> str:
    """Merge recipe family (base / public models: their own family)."""
    if name.startswith("qwen3-0.6b"):
        return "qwen3-0.6b"
    for pre, fam in (("AS3-", "AS3"), ("AV2-", "AV2"), ("V5-", "v5"), ("S5-", "v5"), ("FV-", "FV"), ("Vq-", "Vq"),
                     ("F-", "v1-F/P"),
                     ("P-", "v1-F/P"), ("A-", "A-v3@6750"), ("W-", "W"), ("V-", "v4"), ("S-", "v4")):
        if name.startswith(pre):
            return fam
    return name


def diverse_set(names: list[str], fulls: dict, tasks: list[str], per_family: int) -> tuple[list[str], dict]:
    fam = defaultdict(list)
    for n in names:
        fam[family(n)].append(n)
    keep, info = [], {}
    for f, ms in sorted(fam.items()):
        common = [t for t in tasks if all(t in fulls.get(m, {}) for m in ms)]

        def q(m):
            return float(np.mean([fulls[m][t] for t in common])) if common else 0.0
        ms = sorted(ms, key=q)
        if len(ms) > per_family:
            idx = sorted({0, len(ms) // 2, len(ms) - 1} if per_family >= 3 else {0, len(ms) - 1})
            pick = [ms[i] for i in idx][:per_family]
        else:
            pick = ms
        keep += pick
        info[f] = {"members": ms, "kept": pick}
    return sorted(keep), info


# ----------------------------------------------------------------------------- item models
def is_binary(Y: np.ndarray) -> bool:
    v = Y[~np.isnan(Y)]
    return bool(len(v)) and bool(np.all((v == 0) | (v == 1)))


def fit_linear(Y: np.ndarray, th: np.ndarray) -> dict:
    """Per-item OLS y = a*theta + b (theta standardised), residual variance (floored)."""
    tc = th - th.mean()
    a = (tc[:, None] * (Y - Y.mean(0))).sum(0) / max((tc ** 2).sum(), 1e-12)
    b = Y.mean(0) - a * th.mean()
    r = Y - (a[None] * th[:, None] + b[None])
    s2 = np.maximum((r ** 2).mean(0), 1e-4 + 1e-2 * Y.var(0).mean())
    return {"kind": "linear", "a": a, "b": b, "s2": s2}


def fit_2pl(Y: np.ndarray, th: np.ndarray, l2: float = 1.0, iters: int = 30) -> dict:
    """Per-item logistic y ~ sigmoid(alpha*theta + beta), L2 on (alpha, beta); a = alpha, d = -beta/alpha."""
    N = Y.shape[1]
    al, be = np.zeros(N), np.zeros(N)
    X = np.stack([th, np.ones_like(th)], 1)                       # M x 2
    for _ in range(iters):
        z = th[:, None] * al[None] + be[None]
        p = 1 / (1 + np.exp(-z))
        w = p * (1 - p)
        g0 = (X[:, 0:1] * (Y - p)).sum(0) - l2 * al
        g1 = (X[:, 1:2] * (Y - p)).sum(0) - l2 * be
        h00 = -(w * th[:, None] ** 2).sum(0) - l2
        h01 = -(w * th[:, None]).sum(0)
        h11 = -w.sum(0) - l2
        det = h00 * h11 - h01 ** 2
        da = (h11 * g0 - h01 * g1) / det
        db = (-h01 * g0 + h00 * g1) / det
        al, be = al - da, be - db
    return {"kind": "2pl", "alpha": al, "beta": be}


def item_info(model: dict, th: np.ndarray) -> np.ndarray:
    """Fisher information I_i(theta_m): M x N."""
    if model["kind"] == "linear":
        return np.broadcast_to(model["a"] ** 2 / model["s2"], (len(th), len(model["a"]))).copy()
    p = 1 / (1 + np.exp(-(th[:, None] * model["alpha"][None] + model["beta"][None])))
    return model["alpha"][None] ** 2 * p * (1 - p)


def predict_items(model: dict, theta: float) -> np.ndarray:
    if model["kind"] == "linear":
        return model["a"] * theta + model["b"]
    return 1 / (1 + np.exp(-(theta * model["alpha"] + model["beta"])))


def map_theta(model: dict, y: np.ndarray, cols: np.ndarray) -> float:
    """MAP ability (N(0,1) prior) from the observed selected items."""
    ok = ~np.isnan(y)
    cols, y = cols[ok], y[ok]
    if model["kind"] == "linear":
        a, b, s2 = model["a"][cols], model["b"][cols], model["s2"][cols]
        return float((a * (y - b) / s2).sum() / ((a ** 2 / s2).sum() + 1.0))
    al, be = model["alpha"][cols], model["beta"][cols]
    t = 0.0
    for _ in range(30):
        p = 1 / (1 + np.exp(-(t * al + be)))
        g = (al * (y - p)).sum() - t
        h = -(al ** 2 * p * (1 - p)).sum() - 1
        t -= g / h
    return float(t)


def weights_of(ts: L.TaskScores) -> np.ndarray:
    """Per-item weight in the pool score (1/N; multi-subset: 1/(G * n_subset))."""
    if not ts.multi:
        return np.full(ts.N, 1.0 / ts.N)
    gs = ts.group
    uniq = list(dict.fromkeys(gs.tolist()))
    cnt = {g: int((gs == g).sum()) for g in uniq}
    return np.array([1.0 / (len(uniq) * cnt[g]) for g in gs])


# ----------------------------------------------------------------------------- selection
def select_anchor(Yt: np.ndarray, K: int, u: np.ndarray, seed: int = 0) -> tuple[np.ndarray, np.ndarray]:
    from sklearn.cluster import KMeans
    N = Yt.shape[1]
    if K >= N:
        return np.arange(N), u.copy()
    X = Yt.T                                                          # N x M
    km = KMeans(n_clusters=K, n_init=3, random_state=seed, max_iter=200).fit(X, sample_weight=u)
    sel, w = [], []
    for k in range(K):
        mem = np.where(km.labels_ == k)[0]
        if not len(mem):
            continue
        d = ((X[mem] - km.cluster_centers_[k]) ** 2).sum(1)
        sel.append(int(mem[np.argmin(d)]))
        w.append(float(u[mem].sum()))
    return np.array(sel), np.array(w)


def select_fisher(info: np.ndarray, K: int, u: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
    M, N = info.shape
    if K >= N:
        return np.arange(N), u.copy()
    tot = np.zeros(M)
    chosen = []
    avail = np.ones(N, bool)
    eps = 1e-3 + 1e-3 * info.mean()
    for _ in range(K):
        gain = np.log(eps + tot[:, None] + info).sum(0) - np.log(eps + tot).sum()
        gain[~avail] = -np.inf
        j = int(np.argmax(gain))
        chosen.append(j)
        avail[j] = False
        tot += info[:, j]
    sel = np.array(sorted(chosen))
    return sel, u[sel] / u[sel].sum()


def stratified_random(ts: L.TaskScores, pool: dict, K: int, rng: np.random.RandomState) -> np.ndarray:
    """Stratified random K for set-metric tasks: strata = subset x (gold-score bin | label | class)."""
    N = ts.N
    if K >= N:
        return np.arange(N)
    byid = {x["id"]: x for x in pool["items"]}
    keys = []
    for j, i in enumerate(ts.ids):
        x = byid[i]
        if ts.type == "STS":
            k = str(min(4, int(5 * (x["score"] - pool.get("min_score", 0)) /
                               max(pool.get("max_score", 5) - pool.get("min_score", 0), 1e-9))))
        elif ts.type == "Clustering":
            k = str(x["labels"][0])
        else:
            k = str(x["label"])
        keys.append(f"{x.get('subset', '')}/{k}")
    groups = defaultdict(list)
    for j, k in enumerate(keys):
        groups[k].append(j)
    alloc = {k: len(v) * K / N for k, v in groups.items()}
    base = {k: int(np.floor(a)) for k, a in alloc.items()}
    rest = K - sum(base.values())
    frac = sorted(groups, key=lambda k: -(alloc[k] - base[k]) + rng.rand() * 1e-6)
    for k in frac[:rest]:
        base[k] += 1
    out = []
    for k, v in groups.items():
        out += rng.choice(v, size=min(base[k], len(v)), replace=False).tolist()
    return np.array(sorted(out))


# ----------------------------------------------------------------------------- one fold
def fold(ts: L.TaskScores, pool: dict, f: np.ndarray, train: np.ndarray, test: np.ndarray, K: int,
         n_random: int, methods=METHODS) -> dict:
    """Predictions of the test models' full score for every (method, estimator), random K and the pool."""
    out = {"test": test.tolist()}
    ft = f[train]
    mu, sd = ft.mean(), ft.std() + 1e-12
    th_t = (ft - mu) / sd
    u = weights_of(ts)
    exact = ts.type == "Clustering"

    def cal(s_all):
        a, b = S1.calibrate(s_all[train], ft)
        return (a + b * s_all)[test]
    # pool baseline
    out["pool"] = cal(ts.pool_scores).tolist()
    # random baseline: single draws (metrics averaged later)
    rng = np.random.RandomState(12345)
    rnd = []
    for _ in range(n_random):
        cols = np.sort(rng.choice(ts.N, min(K, ts.N), replace=False))
        rnd.append(cal(ts.subset(cols, exact=exact)).tolist())
    out["random"] = rnd
    if ts.type not in PER_ITEM:
        rng = np.random.RandomState(777)
        draws, sels = [], []
        for r in range(n_random):
            cols = stratified_random(ts, pool, K, rng)
            draws.append(cal(ts.subset(cols, exact=exact)).tolist())
            sels.append(cols.tolist())
        out["strat"] = draws
        out["strat_sel0"] = sels[0]
        # anchor points on the per-item outcomes (STS: per-pair rank agreement), unweighted set metric
        O = S1.item_outcomes(ts)
        O = np.where(np.isnan(O), np.nan_to_num(np.nanmean(O, 0))[None], O)
        sel, _ = select_anchor(O[train], K, weights_of(ts))
        out["anchorset"] = cal(ts.subset(np.sort(sel), exact=exact)).tolist()
        out["anchorset_sel"] = np.sort(sel).tolist()
        return out
    Y = ts.Y.copy()
    colmean = np.nanmean(Y, 0)
    Y = np.where(np.isnan(Y), np.nan_to_num(colmean)[None], Y)
    Yt = Y[train]
    binary = is_binary(Yt)
    model = fit_2pl(Yt, th_t) if binary else fit_linear(Yt, th_t)
    info = item_info(model, th_t)
    pool_est_obs = (Y * u[None]).sum(1)                                  # == the pool score (decomposable)
    for meth in methods:
        sel, w = select_anchor(Yt, K, u) if meth == "anchor" else select_fisher(info, K, u)
        w = w / w.sum()
        est_mean = (Y[:, sel] * w[None]).sum(1)
        est_pirt = np.zeros(ts.M)
        rest = np.ones(ts.N, bool)
        rest[sel] = False
        for m in range(ts.M):
            th = map_theta(model, Y[m, sel], sel)
            pred = predict_items(model, th)
            est_pirt[m] = (u[sel] * Y[m, sel]).sum() + (u[rest] * pred[rest]).sum()
        ests = {"mean": est_mean, "pirt": est_pirt, "gp": 0.5 * (est_mean + est_pirt)}
        for e, v in ests.items():
            out[f"{meth}/{e}"] = cal(v).tolist()
        out[f"{meth}/sel"] = sel.tolist()
        out[f"{meth}/w"] = w.tolist()
    out["binary"] = binary
    out["pool_obs_check"] = float(np.nanmax(np.abs(pool_est_obs - ts.pool_scores))) if not ts.multi else None
    return out


# ----------------------------------------------------------------------------- jobs
_CACHE: dict = {}


def load_task(task: str, names: tuple) -> tuple[L.TaskScores, dict]:
    if (task, names) not in _CACHE:
        _CACHE.clear()
        res, embs = [], []
        for n in names:
            r, e = L.load_scores(n, task)
            res.append(r)
            embs.append(e)
        pool = L.load_pool(task)
        _CACHE[(task, names)] = (L.TaskScores(task, res, embs, pool), pool)
    return _CACHE[(task, names)]


def run_job(suite, scores_dir, task, names, f, train, test, K, n_random, tag):
    L.configure(suite)
    L.SCORES = Path(scores_dir)
    ts, pool = load_task(task, tuple(names))
    t0 = time.time()
    o = fold(ts, pool, np.asarray(f, float), np.asarray(train), np.asarray(test), K, n_random)
    o.update({"task": task, "tag": tag, "seconds": time.time() - t0})
    return o


def main() -> None:
    L.configure(L.suite_from_argv())
    ap = argparse.ArgumentParser()
    ap.add_argument("--suite", choices=["eng", "ara"], default="eng")
    ap.add_argument("--scores-dir", default=str(L.SCORES))
    ap.add_argument("--models", default=None, help="comma list (default: every scored model)")
    ap.add_argument("--exclude", default="", help="comma list of score dirs to ignore")
    ap.add_argument("--tasks", default=None)
    ap.add_argument("--full-csv", default=None, help="name,task,score (default: <lite>/public_full.csv if present)")
    ap.add_argument("--extra-full-csv", default=None, help="another CSV (e.g. our models' Arabic full scores)")
    ap.add_argument("--K", type=int, default=100)
    ap.add_argument("--K-set", type=int, default=None,
                    help="K for set-metric tasks (STS / clustering / pair classification); default = --K")
    ap.add_argument("--set-method", choices=["auto", "strat", "anchorset"], default="anchorset",
                    help="set-metric selection (anchorset beat strat on tau for every set type at K=100 and K=20)")
    ap.add_argument("--per-family", type=int, default=3)
    ap.add_argument("--all-models", action="store_true", help="no family thinning (v1-style reference set)")
    ap.add_argument("--n-random", type=int, default=10)
    ap.add_argument("--choose-by", default="family", choices=["family", "distinct"],
                    help="validation protocol used to choose the per-item method/estimator")
    ap.add_argument("--jobs", type=int, default=1)
    ap.add_argument("--out", default=None)
    ap.add_argument("--report", default=None)
    args = ap.parse_args()
    L.SCORES = Path(args.scores_dir)
    sfx = ("" if args.K > 20 else "_super") + (f"_kset{args.K_set}" if args.K_set and args.K_set != args.K else "")
    out = Path(args.out) if args.out else REPO / "data" / (("mteb_lite" if args.suite == "eng" else "mteb_lite_ar")
                                                           + sfx + "_v2")
    report = Path(args.report) if args.report else L.OUT / f"selection_report_v2{sfx}"
    manifest_name = L.SUITES[args.suite]["manifest"]
    excl = set(filter(None, args.exclude.split(",")))
    names = sorted(args.models.split(",")) if args.models else sorted(
        p.name for p in L.SCORES.iterdir() if p.is_dir() and p.name not in excl)
    tasks = args.tasks.split(",") if args.tasks else [t for t in L.task_names() if L.pool_path(t).exists()]
    types = {t: L.task_type(t) for t in tasks}
    csv_path = args.full_csv or (str(L.LITE / "public_full.csv") if (L.LITE / "public_full.csv").exists() else None)
    fulls = S1.full_scores(names, csv_path) if args.suite == "eng" else ara_full_scores(names, csv_path)
    if args.extra_full_csv:
        for k, v in S1.full_scores([], args.extra_full_csv).items():
            fulls[k].update(v)
    names = [n for n in names if fulls.get(n)]
    ref, fam_info = (names, {}) if args.all_models else diverse_set(names, fulls, tasks, args.per_family)
    fams = defaultdict(list)
    for n in ref:
        fams[family(n)].append(n)
    print(f"[v2] {len(names)} models with full scores -> reference set {len(ref)} in {len(fams)} families", flush=True)

    jobs, info = [], {}
    k_set = args.K_set or args.K
    for t in tasks:
        Kt = args.K if types[t] in PER_ITEM else k_set
        have = [n for n in ref if (L.SCORES / n / f"{t}.json").exists() and t in fulls[n]]
        if len(have) < 4:
            print(f"[v2] {t}: {len(have)} reference models - skipped", flush=True)
            continue
        f = [fulls[n][t] for n in have]
        info[t] = {"names": have, "f": f}
        idx = np.arange(len(have))
        jobs.append((args.suite, str(L.SCORES), t, have, f, idx.tolist(), idx.tolist(), Kt, args.n_random, "final"))
        for m in idx:
            jobs.append((args.suite, str(L.SCORES), t, have, f, [j for j in idx if j != m], [int(m)], Kt,
                         args.n_random, f"distinct:{have[m]}"))
        for fam in sorted({family(n) for n in have}):
            te = [j for j in idx if family(have[j]) == fam]
            tr = [j for j in idx if family(have[j]) != fam]
            if len(tr) >= 4:
                jobs.append((args.suite, str(L.SCORES), t, have, f, tr, te, Kt, args.n_random, f"family:{fam}"))
    print(f"[v2] {len(info)} tasks, {len(jobs)} jobs", flush=True)
    results = []
    if args.jobs > 1:
        from concurrent.futures import ProcessPoolExecutor
        with ProcessPoolExecutor(args.jobs) as ex:
            for r in ex.map(run_job, *zip(*jobs), chunksize=1):
                results.append(r)
    else:
        results = [run_job(*j) for j in jobs]
    by = defaultdict(lambda: defaultdict(dict))        # task -> protocol -> {model idx: {key: pred}}
    final = {}
    for r in results:
        if r["tag"] == "final":
            final[r["task"]] = r
            continue
        proto = r["tag"].split(":")[0]
        for k, v in r.items():
            if k in ("test", "task", "tag", "seconds", "binary", "pool_obs_check") or k.endswith("/sel") \
                    or k.endswith("/w") or k in ("strat_sel0", "anchorset_sel"):
                continue
            for pos, m in enumerate(r["test"]):
                if k in ("random", "strat"):
                    by[r["task"]][proto].setdefault(m, {})[k] = [d[pos] for d in v]
                else:
                    by[r["task"]][proto].setdefault(m, {})[k] = v[pos]
    keys_item = [f"{m}/{e}" for m in METHODS for e in ESTIMATORS]
    rep = {"args": vars(args), "reference": ref, "families": {k: v for k, v in fam_info.items()},
           "per_task": {}, "aggregate": {}, "choice": {}}

    def task_metrics(t, proto, key):
        d = by[t][proto]
        ms = sorted(d)
        tru = np.array([info[t]["f"][m] for m in ms])
        if key in ("random", "strat"):
            nd = len(d[ms[0]][key]) if ms else 0
            per = [S1.metrics(np.array([d[m][key][r] for m in ms]), tru) for r in range(nd)]
            o = {k: float(np.nanmean([p[k] for p in per])) for k in ("kendall", "pearson", "mae")}
            o["n"] = len(ms)
            return o
        return S1.metrics(np.array([d[m][key] for m in ms]), tru)
    # choose one per-item method/estimator for all per-item tasks by the aggregate of the chosen protocol
    for proto in ("family", "distinct"):
        for t in info:
            row = rep["per_task"].setdefault(t, {"type": types[t], "n_ref": len(info[t]["names"])})
            keys = (keys_item if types[t] in PER_ITEM else ["strat", "anchorset"]) + ["random", "pool"]
            row[proto] = {k: task_metrics(t, proto, k) for k in keys if by[t][proto]}

    def agg(proto, key_of):
        """Aggregate over tasks for the models predicted in every task: macro (mean of the type means) and mean,
        x100. Keys holding single random draws are scored per draw and the metrics averaged."""
        models = None
        for t in info:
            ms = {info[t]["names"][m] for m in by[t][proto]}
            models = ms if models is None else models & ms
        models = sorted(models or [])
        if len(models) < 3:
            return {}
        draws = any(key_of(types[t]) in ("random", "strat") for t in info)
        res = []
        for r in range(args.n_random if draws else 1):
            pm, pmean, tm, tmean = [], [], [], []
            for n in models:
                pred, tru = {}, {}
                for t in info:
                    m = info[t]["names"].index(n)
                    v = by[t][proto][m][key_of(types[t])]
                    pred[t] = v[r % len(v)] if isinstance(v, list) else v
                    tru[t] = info[t]["f"][m]
                a_, b_ = S1.aggregate(pred, types), S1.aggregate(tru, types)
                pm.append(a_[0]); pmean.append(a_[1]); tm.append(b_[0]); tmean.append(b_[1])
            res.append({"macro": S1.metrics(np.array(pm), np.array(tm)) | {"tau_resolvable": res_tau(pm, tm)},
                        "mean": S1.metrics(np.array(pmean), np.array(tmean)) | {"tau_resolvable": res_tau(pmean, tmean)},
                        "pred": {"macro": pm, "mean": pmean}})
        out_ = {w: {k: float(np.nanmean([x[w][k] for x in res])) for k in ("kendall", "pearson", "mae", "tau_resolvable")}
                for w in ("macro", "mean")}
        for w in out_:
            out_[w]["n"] = len(models)
        out_["models"] = models
        out_["true"] = {"macro": tm, "mean": tmean}
        out_["pred_draw0"] = res[0]["pred"]
        return out_
    for proto in ("family", "distinct"):
        A = {}
        for k in keys_item:
            for sm in ("strat", "anchorset"):
                A[f"{k}+{sm}"] = agg(proto, lambda ty, k=k, sm=sm: k if ty in PER_ITEM else sm)
        A["random"] = agg(proto, lambda ty: "random")
        A["pool"] = agg(proto, lambda ty: "pool")
        rep["aggregate"][proto] = A
    cands = {k: v for k, v in rep["aggregate"][args.choose_by].items() if "+" in k and v
             and (args.set_method == "auto" or k.endswith("+" + args.set_method))}
    best = (max(cands, key=lambda k: (cands[k]["macro"]["kendall"], -cands[k]["macro"]["mae"])) if cands
            else "anchor/mean+strat")
    best_item, best_set = best.split("+")
    rep["choice"] = {"combo": best, "per_item": best_item, "set_metric": best_set, "chosen_by": args.choose_by}
    meth, est = best_item.split("/")

    # final selection on the whole reference set
    (out / "selection").mkdir(parents=True, exist_ok=True)
    man = {"benchmark": f"MTEB-lite v2 ({args.suite})", "K": args.K, "K_set": args.K_set or args.K, "reference_models": ref,
           "method": {"per_item": best_item, "set_metric": f"{best_set} + calibration",
                      "item_model": "linear-Gaussian (continuous) / 2PL (binary)"}, "tasks": {}}
    for t, r in final.items():
        ts, pool = load_task(t, tuple(info[t]["names"]))
        f = np.array(info[t]["f"])
        doc = {"task": t, "type": types[t], "K": None, "reference_models": info[t]["names"]}
        if types[t] in PER_ITEM:
            sel, w = np.array(r[f"{meth}/sel"]), np.array(r[f"{meth}/w"])
            Y = np.where(np.isnan(ts.Y), np.nan_to_num(np.nanmean(ts.Y, 0))[None], ts.Y)
            th = (f - f.mean()) / (f.std() + 1e-12)
            model = fit_2pl(Y, th) if r["binary"] else fit_linear(Y, th)
            est_vals = {"mean": (Y[:, sel] * w[None]).sum(1)}
            u = weights_of(ts)
            rest = np.ones(ts.N, bool)
            rest[sel] = False
            pir = []
            for m in range(ts.M):
                tm_ = map_theta(model, Y[m, sel], sel)
                pir.append((u[sel] * Y[m, sel]).sum() + (u[rest] * predict_items(model, tm_)[rest]).sum())
            est_vals["pirt"] = np.array(pir)
            est_vals["gp"] = 0.5 * (est_vals["mean"] + est_vals["pirt"])
            a, b = S1.calibrate(est_vals[est], f)
            params = ({"kind": "linear", "a": model["a"].tolist(), "b": model["b"].tolist(), "s2": model["s2"].tolist()}
                      if model["kind"] == "linear" else
                      {"kind": "2pl", "alpha": model["alpha"].tolist(), "beta": model["beta"].tolist()})
            doc.update({"ids": [ts.ids[j] for j in sel], "weights": w.tolist(), "estimator": est,
                        "theta_scale": {"mu": float(f.mean()), "sd": float(f.std() + 1e-12)},
                        "pool_ids": ts.ids, "pool_weights": u.tolist(), "item_model": params,
                        "calibration": {"a": a, "b": b}})
        else:
            sel = np.array(r["strat_sel0"] if best_set == "strat" else r["anchorset_sel"])
            exact = types[t] == "Clustering"
            a, b = S1.calibrate(ts.subset(sel, exact=exact), f)
            doc.update({"ids": [ts.ids[j] for j in sel], "estimator": "set_metric", "set_selection": best_set,
                        "calibration": {"a": a, "b": b}})
        doc["K"] = len(doc["ids"])
        (out / "selection" / f"{t}.json").write_text(json.dumps(doc))
        man["tasks"][t] = {"type": types[t], "main_score": L.get_task(t).metadata.main_score,
                           "selection": f"selection/{t}.json", "K": doc["K"], "estimator": doc["estimator"]}
    (out / manifest_name).write_text(json.dumps(man, indent=1))
    report.parent.mkdir(parents=True, exist_ok=True)
    report.with_suffix(".json").write_text(json.dumps(rep, indent=1, default=float))
    md = render(rep, args, best, fams)
    report.with_suffix(".md").write_text(md)
    print(md)


def res_tau(pred, true, thr: float = 0.5) -> float:
    """Kendall-style agreement on the 'resolvable' pairs only (true scores differ by >= thr points)."""
    p, t = np.asarray(pred, float), np.asarray(true, float)
    c = d = 0
    for i in range(len(t)):
        for j in range(i + 1, len(t)):
            if abs(t[i] - t[j]) >= thr:
                if np.sign(p[i] - p[j]) == np.sign(t[i] - t[j]):
                    c += 1
                else:
                    d += 1
    return (c - d) / (c + d) if c + d else float("nan")


def ara_full_scores(names: list[str], csv_path) -> dict:
    """Arabic: our models from <lite>/ours_full.csv (mteb_lite_ara_full.py) + public CSV."""
    fulls = defaultdict(dict)
    for p in [L.LITE / "ours_full.csv", csv_path]:
        if p and Path(p).exists():
            for k, v in S1.full_scores([], str(p)).items():
                fulls[k].update(v)
    return fulls


def render(rep: dict, args, best: str, fams: dict) -> str:
    def f3(x):
        return "-" if x is None or (isinstance(x, float) and np.isnan(x)) else f"{x:.3f}"
    L_ = [f"# MTEB-lite v2 selection report ({args.suite}, K={args.K})", "",
          f"Reference set ({len(rep['reference'])} models, {len(fams)} families, <= {args.per_family} per family): "
          + "; ".join(f"**{k}**: {', '.join(v)}" for k, v in sorted(fams.items())), "",
          f"K={args.K} for per-item tasks, K={args.K_set or args.K} for set-metric tasks. "
          f"Chosen: **{best}** (per-item method/estimator by the {rep['choice']['chosen_by']}"
          f"-level LOMO macro tau among {{anchor, fisher}} x {{mean, pirt, gp}}; set-metric selection: {args.set_method}). Set-metric "
          "tasks (STS, clustering, pair classification): no per-item IRT (the metric is not a sum over items) -> "
          "stratified random K ('strat') or unweighted anchor points on per-item outcomes ('anchorset'), + "
          f"calibration. Random = uniform K, metrics averaged over {args.n_random} single draws. tau_res = tau on "
          "the pairs whose true score differs by >= 0.5 points.", ""]
    for proto in ("family", "distinct"):
        A = rep["aggregate"].get(proto, {})
        if not A:
            continue
        n = next((v["macro"]["n"] for v in A.values() if v), 0)
        L_ += [f"## Aggregate, {proto}-level LOMO ({n} held-out models with all tasks; scores x100)", "",
               "| estimate | macro tau | macro tau_res | macro r | macro MAE | mean tau | mean tau_res | mean r | mean MAE |",
               "|---|---|---|---|---|---|---|---|---|"]
        for k, v in A.items():
            if v:
                L_.append(f"| {k}{' (chosen)' if k == best else ''} | {f3(v['macro']['kendall'])} | "
                          f"{f3(v['macro']['tau_resolvable'])} | {f3(v['macro']['pearson'])} | "
                          f"{f3(v['macro']['mae'])} | {f3(v['mean']['kendall'])} | {f3(v['mean']['tau_resolvable'])} | "
                          f"{f3(v['mean']['pearson'])} | {f3(v['mean']['mae'])} |")
        L_.append("")
    L_ += ["## Per task (family-level LOMO; chosen estimator vs random K vs pool)", "",
           "| task | type | ref | chosen tau | chosen r | chosen MAE | random tau | random MAE | pool tau | pool MAE |",
           "|---|---|---|---|---|---|---|---|---|---|"]
    for t, r in rep["per_task"].items():
        fam = r.get("family", {})
        k = best.split("+")[0] if r["type"] in PER_ITEM else best.split("+")[1]
        c, ra, po = fam.get(k, {}), fam.get("random", {}), fam.get("pool", {})
        L_.append(f"| {t} | {r['type']} | {r['n_ref']} | {f3(c.get('kendall'))} | {f3(c.get('pearson'))} | "
                  f"{f3(c.get('mae'))} | {f3(ra.get('kendall'))} | {f3(ra.get('mae'))} | {f3(po.get('kendall'))} | "
                  f"{f3(po.get('mae'))} |")
    return "\n".join(L_) + "\n"


if __name__ == "__main__":
    main()