Blame for code/mteb_lite_lib.py

"""MTEB-lite: shared code for the pool builder, the per-item scorer, the selection and the lite eval.

Two suites (configure(suite) / --suite):
  eng  MTEB(eng, v2), 41 tasks            data/mteb_lite/pool, outputs/mteb_lite/scores
  ara  our 42 Arabic MMTEB test tasks     data/mteb_lite_ar/pool, outputs/mteb_lite_ar/scores
       (configs/merge/mmteb_ar_test_tasks.txt; each restricted to its Arabic hf_subsets and test split exactly
       as run_mteb_arabic.py --suite test: adip.eval.mteb_arabic.suite_specs / get_mteb_task)

Pool formats (<pool dir>/<task>.json; common keys: task, type, main_score, eval_split, hf_subset). Items of
multi-subset tasks carry "subset" (retrieval: query_subset {qid: subset}); task scores are means over subsets,
as in mteb:
  Retrieval       queries {qid: text}, corpus {did: text}, qrels {qid: {did: rel}}, candidates {qid: [did]},
                  random_docs [did] (multi-subset: random_docs_by_subset {subset: [did]}), ignore_identical_ids
  Reranking       items [{id, query (str | list[str]), positive [str], negative [str], (candidates [str])}]
  Classification  train {text, label, orig_index}, experiments [[train idx]], items [{id, text, label}]
                  (multi-subset: groups {subset: {train, experiments}})
  PairClassification items [{id, sentence1, sentence2, label}]
  STS             items [{id, sentence1, sentence2, score}]
  Clustering      items [{id, text, labels [level0, level1, ...]}], mteb_ids (mteb's own downsample, if any)
  Summarization   items [{id, machine [str], human [str], relevance [float]}]
  BitextMining    items [{id, subset, sentence1, target}], bitext_texts {key: text},
                  bitext_candidates {subset: [key]} (fixed distractors; an item is correct if its own target
                  is the nearest of distractors + target)

Per-item scores (<scores dir>/<NAME>/<task>.json): {"task", "type", "pool_score", "pool_metrics", "item_key",
"items": [{id, ...}]}; the item field used by the selection is item_key (default ITEM_KEY[type]). Clustering
also writes the pool embeddings (<task>.emb.npy, float16, row order = pool items) so subset scores can
re-cluster exactly as the lite eval does.
"""
from __future__ import annotations

import json
import math
import random
import sys
from collections import defaultdict
from pathlib import Path
from typing import Any, Callable, Optional, Sequence

import numpy as np

REPO = Path(__file__).resolve().parents[1]
SUITE = "eng"
LITE = REPO / "data" / "mteb_lite"
POOL = LITE / "pool"
SCORES = REPO / "outputs" / "mteb_lite" / "scores"
OUT = REPO / "outputs" / "mteb_lite"
BENCHMARK = "MTEB(eng, v2)"
AR_TASKS = REPO / "configs" / "merge" / "mmteb_ar_test_tasks.txt"
SUITES = {
    "eng": {"lite": REPO / "data" / "mteb_lite", "out": REPO / "outputs" / "mteb_lite", "manifest": "MTEB_lite_eng.json",
            "max_length": None},
    "ara": {"lite": REPO / "data" / "mteb_lite_ar", "out": REPO / "outputs" / "mteb_lite_ar",
            "manifest": "MTEB_lite_ara.json", "max_length": 512},     # run_mteb_arabic.py --max-length default
}

# per-item field used for selection; "set" types are scored on item sets by TaskScores.subset
ITEM_KEY = {"Retrieval": "ndcg10", "Reranking": "ap", "Classification": "acc", "PairClassification": "correct",
            "STS": "cos", "Clustering": "agree", "Summarization": "spearman", "BitextMining": "correct"}
SET_TYPES = {"STS", "Clustering", "PairClassification"}
KMEANS_RUNS = 3


def configure(suite: str) -> None:
    """Point the module at a suite's pools / scores (module globals are read at call time)."""
    global SUITE, LITE, POOL, SCORES, OUT
    if suite not in SUITES:
        raise ValueError(f"unknown suite {suite!r} (eng | ara)")
    SUITE = suite
    LITE = SUITES[suite]["lite"]
    POOL = LITE / "pool"
    OUT = SUITES[suite]["out"]
    SCORES = OUT / "scores"
    _TASKS.clear()


def suite_from_argv(argv: Optional[list[str]] = None) -> str:
    """--suite from the command line (before the full parser is built, so defaults follow the suite)."""
    argv = sys.argv[1:] if argv is None else argv
    for i, a in enumerate(argv):
        if a == "--suite" and i + 1 < len(argv):
            return argv[i + 1]
        if a.startswith("--suite="):
            return a.split("=", 1)[1]
    return "eng"


# ----------------------------------------------------------------------------- tasks / pools
_TASKS: dict = {}


def _ar_specs() -> dict:
    sys.path.insert(0, str(REPO / "src"))
    from adip.eval import mteb_arabic as ma
    names = [l.strip() for l in AR_TASKS.read_text().splitlines() if l.strip()]
    return {s["task"]: s for s in ma.suite_specs("test", tasks=names)}


def benchmark_tasks():
    """{name: mteb task} of the current suite (ara: restricted to the Arabic hf_subsets, with the split the
    Arabic eval uses in task._lite_split)."""
    if not _TASKS:
        if SUITE == "eng":
            import mteb
            for t in mteb.get_benchmark(BENCHMARK).tasks:
                t._lite_split = t.eval_splits[0]
                _TASKS[t.metadata.name] = t
        else:
            specs = _ar_specs()                   # also puts src/ on sys.path
            from adip.eval import mteb_arabic as ma
            for n, s in specs.items():
                t = ma.get_mteb_task(s)
                t._lite_split = s["split"]
                t._lite_spec = s
                _TASKS[n] = t
    return _TASKS


def task_names() -> list[str]:
    if SUITE == "ara":
        return [l.strip() for l in AR_TASKS.read_text().splitlines() if l.strip()]
    return list(benchmark_tasks())


def get_task(name: str):
    return benchmark_tasks()[name]


def task_type(name: str) -> str:
    return get_task(name).metadata.type


def pool_path(name: str) -> Path:
    return POOL / f"{name}.json"


def save_pool(pool: dict) -> None:
    POOL.mkdir(parents=True, exist_ok=True)
    p = pool_path(pool["task"])
    tmp = p.with_suffix(".tmp")
    tmp.write_text(json.dumps(pool, ensure_ascii=False))
    tmp.replace(p)


def load_pool(name: str) -> dict:
    return json.loads(pool_path(name).read_text())


def doc_text(v: Any) -> str:
    """mteb's corpus_to_str: title + ' ' + text, stripped."""
    if isinstance(v, str):
        return v.strip()
    if "title" in v and v.get("title"):
        return (v["title"] + " " + v["text"]).strip()
    return v["text"].strip()


def item_ids(pool: dict) -> list[str]:
    if pool["type"] == "Retrieval":
        return list(pool["queries"])
    return [x["id"] for x in pool["items"]]


def item_groups(pool: dict) -> dict[str, str]:
    """{item id: hf_subset} ('' for single-subset tasks)."""
    if pool["type"] == "Retrieval":
        qs = pool.get("query_subset", {})
        return {q: qs.get(q, "") for q in pool["queries"]}
    return {x["id"]: x.get("subset", "") for x in pool["items"]}


def pool_size_str(pool: dict) -> str:
    t = pool["type"]
    if t == "Retrieval":
        return (f"{len(pool['queries'])} queries, {len(pool['corpus'])} docs "
                f"(full {pool['full_sizes']['queries']} q / {pool['full_sizes']['corpus']} docs)")
    if t == "Reranking":
        n = sum(len(x.get("candidates") or (x["positive"] + x["negative"])) for x in pool["items"])
        return f"{len(pool['items'])} queries, {n} candidates (full {pool['full_sizes']['queries']} q)"
    if t == "Classification":
        ntr = (sum(len(g["train"]["text"]) for g in pool["groups"].values()) if "groups" in pool
               else len(pool["train"]["text"]))
        return f"{len(pool['items'])} test, {ntr} train texts (full {pool['full_sizes']})"
    if t == "Summarization":
        return f"{len(pool['items'])} documents"
    if t == "BitextMining":
        return (f"{len(pool['items'])} pairs, {len(pool['bitext_texts'])} target texts "
                f"({len(pool['bitext_candidates'])} subsets; full {pool['full_sizes']})")
    return f"{len(pool['items'])} items (full {pool['full_sizes']})"


def texts_to_encode(pool: dict) -> list[tuple[str, Optional[str]]]:
    """(text, role) pairs a model has to encode for this pool; role: query | document | None."""
    t = pool["type"]
    if t == "Retrieval":
        return [(q, "query") for q in pool["queries"].values()] + [(d, "document") for d in pool["corpus"].values()]
    if t == "Reranking":
        out = []
        for x in pool["items"]:
            qs = x["query"] if isinstance(x["query"], list) else [x["query"]]
            docs = x.get("candidates") or (x["positive"] + x["negative"])
            out += [(q, "query") for q in qs] + [(d, "document") for d in docs]
        return out
    if t == "Classification":
        trs = ([s for g in pool["groups"].values() for s in g["train"]["text"]] if "groups" in pool
               else pool["train"]["text"])
        return [(s, None) for s in trs] + [(x["text"], None) for x in pool["items"]]
    if t in ("PairClassification", "STS"):
        return [(x[k], None) for x in pool["items"] for k in ("sentence1", "sentence2")]
    if t == "Clustering":
        return [(x["text"], None) for x in pool["items"]]
    if t == "Summarization":
        return [(s, None) for x in pool["items"] for s in x["machine"] + x["human"]]
    if t == "BitextMining":
        return [(x["sentence1"], None) for x in pool["items"]] + [(v, None) for v in pool["bitext_texts"].values()]
    raise ValueError(t)


def restrict_pool(pool: dict, ids: Sequence[str]) -> dict:
    """The pool restricted to the given items (what the lite eval encodes)."""
    keep = set(ids)
    p = dict(pool)
    if pool["type"] == "Retrieval":
        qs = [q for q in pool["queries"] if q in keep]
        if "random_docs_by_subset" in pool:
            subs = {pool["query_subset"][q] for q in qs}
            docs = {d for sub in subs for d in pool["random_docs_by_subset"][sub]}
            p["random_docs_by_subset"] = {sub: pool["random_docs_by_subset"][sub] for sub in subs}
            p["query_subset"] = {q: pool["query_subset"][q] for q in qs}
        else:
            docs = set(pool["random_docs"])
        for q in qs:
            docs.update(pool["candidates"][q])
        p["queries"] = {q: pool["queries"][q] for q in qs}
        p["candidates"] = {q: pool["candidates"][q] for q in qs}
        p["qrels"] = {q: pool["qrels"][q] for q in qs}
        p["corpus"] = {d: v for d, v in pool["corpus"].items() if d in docs}
    else:
        p["items"] = [x for x in pool["items"] if x["id"] in keep]
        if "mteb_ids" in p:
            p["mteb_ids"] = [i for i in p["mteb_ids"] if i in keep]
        if pool["type"] == "BitextMining":
            subs = {x["subset"] for x in p["items"]}
            p["bitext_candidates"] = {k: v for k, v in pool["bitext_candidates"].items() if k in subs}
            need = {k for v in p["bitext_candidates"].values() for k in v} | {x["target"] for x in p["items"]}
            p["bitext_texts"] = {k: v for k, v in pool["bitext_texts"].items() if k in need}
    return p


def group_mean(values: np.ndarray, groups: Sequence[str]) -> float:
    """Mean over hf_subsets of the per-subset nanmean (mteb's aggregation); plain nanmean for one subset."""
    values = np.asarray(values, dtype=float)
    gs = np.asarray(list(groups))
    uniq = list(dict.fromkeys(gs.tolist()))
    if len(uniq) <= 1:
        return float(np.nanmean(values)) if len(values) else float("nan")
    with np.errstate(all="ignore"):
        per = [np.nanmean(values[gs == g]) for g in uniq]
    return float(np.nanmean(per))


# ----------------------------------------------------------------------------- encoding
def encode_texts(model, texts: list[str], task_name: str, role: Optional[str], batch_size: int = 64) -> np.ndarray:
    """model.encode exactly as mteb's evaluators call it (task_name + prompt_type); dedupes texts;
    halves the batch on CUDA OOM."""
    from mteb.encoder_interface import PromptType
    pt = {"query": PromptType.query, "document": PromptType.document, None: None}[role]
    uniq = list(dict.fromkeys(texts))
    if not uniq:
        return np.zeros((0, 1), dtype=np.float32)
    bs = batch_size
    while True:
        try:
            emb = model.encode(uniq, task_name=task_name, prompt_type=pt, batch_size=bs)
            break
        except Exception as e:  # torch.cuda.OutOfMemoryError
            if "out of memory" not in str(e).lower() or bs == 1:
                raise
            import torch
            torch.cuda.empty_cache()
            bs = max(1, bs // 2)
            print(f"[encode] OOM -> batch {bs}", flush=True)
    if hasattr(emb, "detach"):
        emb = emb.detach().cpu().float().numpy()
    emb = np.asarray(emb, dtype=np.float32)
    pos = {s: i for i, s in enumerate(uniq)}
    return emb[[pos[s] for s in texts]]


def similarity_name(model) -> str:
    meta = getattr(model, "mteb_model_meta", None) or getattr(model, "model_meta", None)
    name = getattr(meta, "similarity_fn_name", None) or "cosine"
    return str(getattr(name, "value", name)).lower()


def sim_matrix(a: np.ndarray, b: np.ndarray, sim: str = "cosine") -> np.ndarray:
    if sim.startswith("cos"):
        a = a / np.maximum(np.linalg.norm(a, axis=1, keepdims=True), 1e-12)
        b = b / np.maximum(np.linalg.norm(b, axis=1, keepdims=True), 1e-12)
    return a @ b.T


def cos_pairs(a: np.ndarray, b: np.ndarray) -> np.ndarray:
    a = a / np.maximum(np.linalg.norm(a, axis=1, keepdims=True), 1e-12)
    b = b / np.maximum(np.linalg.norm(b, axis=1, keepdims=True), 1e-12)
    return (a * b).sum(1)


# ----------------------------------------------------------------------------- metrics
def ndcg_at_k(ranked: Sequence[str], qrel: dict, k: int = 10) -> float:
    """pytrec_eval ndcg_cut (gain = relevance grade, log2 discount)."""
    dcg = sum(max(qrel.get(d, 0), 0) / math.log2(i + 2) for i, d in enumerate(ranked[:k]))
    ideal = sorted((v for v in qrel.values() if v > 0), reverse=True)[:k]
    idcg = sum(v / math.log2(i + 2) for i, v in enumerate(ideal))
    return dcg / idcg if idcg > 0 else 0.0


def recall_at_k(ranked: Sequence[str], qrel: dict, k: int = 10) -> float:
    rel = {d for d, v in qrel.items() if v > 0}
    return len(rel & set(ranked[:k])) / len(rel) if rel else 0.0


def average_precision(y: np.ndarray, s: np.ndarray, w: Optional[np.ndarray] = None) -> float:
    from sklearn.metrics import average_precision_score
    y = np.asarray(y)
    if y.sum() == 0 or y.sum() == len(y):
        return float("nan")
    return float(average_precision_score(y, s, sample_weight=w))


def pair_weights(pool: dict, labels: np.ndarray) -> Optional[np.ndarray]:
    lw = pool.get("label_weights")
    return None if not lw else np.array([lw[str(int(v))] for v in labels], dtype=float)


def spearman(x, y) -> float:
    from scipy.stats import spearmanr
    x, y = np.asarray(x, float), np.asarray(y, float)
    if len(x) < 3 or np.all(x == x[0]) or np.all(y == y[0]):
        return float("nan")
    return float(spearmanr(x, y).statistic if hasattr(spearmanr(x, y), "statistic") else spearmanr(x, y)[0])


def v_measure_fast(labels: np.ndarray, clusters: np.ndarray) -> float:
    """sklearn's v_measure_score (beta=1) with numpy only (labels / clusters: non-negative ints)."""
    n = len(labels)
    if n == 0:
        return float("nan")
    _, li = np.unique(labels, return_inverse=True)
    _, ci = np.unique(clusters, return_inverse=True)
    nl, nc = li.max() + 1, ci.max() + 1
    if nl == 1 or nc == 1:
        return 1.0 if nl == nc else 0.0
    cont = np.bincount(li * nc + ci, minlength=nl * nc).reshape(nl, nc).astype(float)
    pl, pc, pj = cont.sum(1) / n, cont.sum(0) / n, cont / n
    h_l = -(pl * np.log(pl)).sum()
    h_c = -(pc * np.log(pc)).sum()
    nz = pj > 0
    mi = (pj[nz] * np.log(pj[nz] / (pl[:, None] * pc[None, :])[nz])).sum()
    hom, com = mi / h_l, mi / h_c
    return float(2 * hom * com / (hom + com)) if hom + com > 0 else 0.0


def level_labels(items: list[dict], level: int) -> tuple[np.ndarray, np.ndarray]:
    """(valid mask, int labels) of one hierarchy level."""
    lab = [x["labels"][level] if len(x["labels"]) > level else None for x in items]
    mask = np.array([v is not None for v in lab])
    vocab = {v: i for i, v in enumerate(sorted({v for v in lab if v is not None}))}
    return mask, np.array([vocab[v] for v in lab if v is not None], dtype=int)


def _group_index(items: list[dict]) -> list[np.ndarray]:
    gs = np.array([x.get("subset", "") for x in items])
    return [np.where(gs == g)[0] for g in dict.fromkeys(gs.tolist())]


def kmeans_runs(emb: np.ndarray, items: list[dict], n_levels: int, runs: int = KMEANS_RUNS) -> list[list[np.ndarray]]:
    """Cluster ids [run][level] (k = number of gold labels at that level, -1 for items without the level);
    k-means runs separately inside every hf_subset (cluster ids are only comparable within a subset)."""
    from sklearn.cluster import KMeans
    x = emb / np.maximum(np.linalg.norm(emb, axis=1, keepdims=True), 1e-12)
    groups = _group_index(items)
    out = []
    for r in range(runs):
        per = []
        for lv in range(n_levels):
            ids = np.full(len(items), -1, dtype=int)
            for gi in groups:
                its = [items[i] for i in gi]
                mask, lab = level_labels(its, lv)
                k = len(np.unique(lab))
                gids = np.full(len(gi), -1, dtype=int)
                if mask.sum() >= 2 and k >= 2:
                    km = KMeans(n_clusters=min(k, int(mask.sum())), n_init=1, random_state=r, max_iter=100)
                    gids[mask] = km.fit_predict(x[gi][mask])
                elif mask.any():
                    gids[mask] = 0
                ids[gi] = gids
            per.append(ids)
        out.append(per)
    return out


def cluster_score(emb: np.ndarray, items: list[dict], n_levels: int, runs: int = KMEANS_RUNS) -> float:
    """Lite clustering score of a set of items: mean over k-means runs and levels of the V-measure of
    k-means (k = #gold labels present) on exactly these items (per hf_subset, then averaged over subsets).
    Used by the scorer (pool), the selection (subsets, from stored embeddings) and the lite eval alike."""
    cl = kmeans_runs(emb, items, n_levels, runs)
    per_group = []
    for gi in _group_index(items):
        its = [items[i] for i in gi]
        vals = []
        for per in cl:
            for lv in range(n_levels):
                mask, lab = level_labels(its, lv)
                if mask.sum() >= 2:
                    vals.append(v_measure_fast(lab, per[lv][gi][mask]))
        if vals:
            per_group.append(float(np.nanmean(vals)))
    return float(np.nanmean(per_group)) if per_group else float("nan")


# ----------------------------------------------------------------------------- per-task scoring
def score_pool(pool: dict, model, batch_size: int = 64, log: Callable = print) -> tuple[dict, Optional[np.ndarray]]:
    """Encode the pool with the model (mteb's prompts for this task) and compute per-item + pool scores.
    Returns (result json, embeddings to save or None). Pool scores of multi-subset tasks are means over
    hf_subsets (mteb's aggregation)."""
    t, name = pool["type"], pool["task"]
    sim = similarity_name(model)
    enc = lambda texts, role: encode_texts(model, texts, name, role, batch_size)  # noqa: E731
    emb_out = None
    grp = item_groups(pool)
    item_key = ITEM_KEY[t]
    if t == "Retrieval":
        qids, dids = list(pool["queries"]), list(pool["corpus"])
        Q = enc([pool["queries"][q] for q in qids], "query")
        D = enc([pool["corpus"][d] for d in dids], "document")
        dpos = {d: i for i, d in enumerate(dids)}
        multi = "random_docs_by_subset" in pool
        if multi:       # docs of a subset: its namespaced ids "<subset>|<doc id>"
            dsub = np.array([d.split("|", 1)[0] for d in dids])
            rnd_by = {k: [dpos[d] for d in v] for k, v in pool["random_docs_by_subset"].items()}
        else:
            rnd = [dpos[d] for d in pool["random_docs"]]
        ign = pool.get("ignore_identical_ids", False)
        items, pool_ndcg, pool_rec = [], [], []
        for s in range(0, len(qids), 256):
            S = sim_matrix(Q[s:s + 256], D, sim)
            for r, q in enumerate(qids[s:s + 256]):
                row = S[r].copy()
                if multi:
                    row[dsub != pool["query_subset"][q]] = -np.inf
                if ign and q in dpos:
                    row[dpos[q]] = -np.inf
                top = np.argpartition(-row, min(10, len(row) - 1))[:11]
                top = top[np.argsort(-row[top])]
                ranked = [dids[i] for i in top if not (ign and dids[i] == q) and np.isfinite(row[i])][:10]
                qrel = pool["qrels"][q]
                pool_ndcg.append(ndcg_at_k(ranked, qrel))
                pool_rec.append(recall_at_k(ranked, qrel))
                rr = rnd_by[pool["query_subset"][q]] if multi else rnd
                cand = sorted(set(dpos[d] for d in pool["candidates"][q]) | set(rr))
                if ign:
                    cand = [i for i in cand if dids[i] != q]
                cs = S[r, cand]
                order = np.argsort(-cs)[:10]
                lr = [dids[cand[i]] for i in order]
                items.append({"id": q, "ndcg10": ndcg_at_k(lr, qrel), "recall10": recall_at_k(lr, qrel),
                              "ndcg10_pool": pool_ndcg[-1], "n_cand": len(cand)})
        g = [grp[q] for q in qids]
        res = {"pool_score": group_mean(pool_ndcg, g),
               "pool_metrics": {"ndcg_at_10_pool_corpus": group_mean(pool_ndcg, g),
                                "recall_at_10_pool_corpus": group_mean(pool_rec, g),
                                "ndcg_at_10_local": group_mean([x["ndcg10"] for x in items], g),
                                "recall_at_10_local": group_mean([x["recall10"] for x in items], g)}}
    elif t == "Reranking":
        qtexts, qown, dtexts, down = [], [], [], []
        for i, x in enumerate(pool["items"]):
            qs = x["query"] if isinstance(x["query"], list) else [x["query"]]
            qtexts += qs
            qown += [i] * len(qs)
            ds = x.get("candidates") or (x["positive"] + x["negative"])
            dtexts += ds
            down += [i] * len(ds)
        Q, D = enc(qtexts, "query"), enc(dtexts, "document")
        qown, down = np.array(qown), np.array(down)
        items = []
        for i, x in enumerate(pool["items"]):
            S = sim_matrix(Q[qown == i], D[down == i], sim).max(0)
            if x.get("candidates"):            # MIRACL format: relevance = candidate is a positive
                pos = set(x["positive"])
                y = np.array([int(d in pos) for d in x["candidates"]])
            else:
                y = np.array([1] * len(x["positive"]) + [0] * len(x["negative"]))
            order = np.argsort(-S)
            first = int(np.argmax(y[order])) + 1 if y.any() else 10 ** 9
            items.append({"id": x["id"], "ap": average_precision(y, S),
                          "mrr10": 1.0 / first if first <= 10 else 0.0,
                          "ndcg10": ndcg_at_k([str(j) for j in order], {str(j): int(v) for j, v in enumerate(y)})})
        if "ndcg" in str(pool["main_score"]).lower():
            item_key = "ndcg10"
        g = [x.get("subset", "") for x in pool["items"]]
        res = {"pool_score": group_mean([x[item_key] for x in items], g),
               "pool_metrics": {"map": group_mean([x["ap"] for x in items], g),
                                "mrr_at_10": group_mean([x["mrr10"] for x in items], g),
                                "ndcg_at_10": group_mean([x["ndcg10"] for x in items], g)}}
    elif t == "Classification":
        from sklearn.linear_model import LogisticRegression
        groups = pool["groups"] if "groups" in pool else {"": {"train": pool["train"],
                                                                "experiments": pool["experiments"]}}
        Xte = enc([x["text"] for x in pool["items"]], None)
        yte = np.array([str(x["label"]) for x in pool["items"]])
        gte = np.array([x.get("subset", "") for x in pool["items"]])
        n_exp = max(len(g["experiments"]) for g in groups.values())
        correct = np.full((n_exp, len(yte)), np.nan)
        per_exp = defaultdict(list)
        for sub, gd in groups.items():
            cols = np.where(gte == sub)[0] if "groups" in pool else np.arange(len(yte))
            if not len(cols):
                continue
            Xtr = enc(gd["train"]["text"], None)
            ytr = np.array([str(v) for v in gd["train"]["label"]])
            for e, idx in enumerate(gd["experiments"]):
                clf = LogisticRegression(random_state=42, n_jobs=-1, max_iter=100)
                clf.fit(Xtr[idx], ytr[idx])
                correct[e, cols] = clf.predict(Xte[cols]) == yte[cols]
                per_exp[e].append(float(np.mean(correct[e, cols])))
        acc = np.nanmean(correct, 0)
        items = [{"id": x["id"], "acc": float(a)} for x, a in zip(pool["items"], acc)]
        sc = group_mean(acc, gte)
        res = {"pool_score": sc,
               "pool_metrics": {"accuracy": sc,
                                "accuracy_per_experiment": [float(np.mean(per_exp[e])) for e in sorted(per_exp)]}}
    elif t == "PairClassification":
        A = enc([x["sentence1"] for x in pool["items"]], None)
        B = enc([x["sentence2"] for x in pool["items"]], None)
        y = np.array([x["label"] for x in pool["items"]])
        cos = cos_pairs(A, B)
        scores = {"cosine": cos, "dot": (A * B).sum(1), "euclidean": -np.linalg.norm(A - B, axis=1),
                  "manhattan": -np.abs(A - B).sum(1)}
        w = pair_weights(pool, y)
        gs = np.array([x.get("subset", "") for x in pool["items"]])

        def gap(v, wt):
            return float(np.nanmean([average_precision(y[gs == g], v[gs == g], None if wt is None else wt[gs == g])
                                     for g in dict.fromkeys(gs.tolist())]))
        aps = {k: gap(v, w) for k, v in scores.items()}
        aps_bal = {f"{k}_ap_balanced": gap(v, None) for k, v in scores.items()}
        thr, best = best_accuracy_threshold(cos, y)
        items = [{"id": x["id"], "cos": float(c), "correct": int((c >= thr) == bool(l))}
                 for x, c, l in zip(pool["items"], cos, y)]
        res = {"pool_score": float(np.nanmax(list(aps.values()))),
               "pool_metrics": {**{f"{k}_ap": v for k, v in aps.items()}, **aps_bal, "cosine_accuracy": best,
                                "cosine_threshold": thr}}
    elif t == "STS":
        A = enc([x["sentence1"] for x in pool["items"]], None)
        B = enc([x["sentence2"] for x in pool["items"]], None)
        cos = cos_pairs(A, B)
        items = [{"id": x["id"], "cos": float(c), "gold": x["score"]} for x, c in zip(pool["items"], cos)]
        gold = np.array([x["score"] for x in pool["items"]], dtype=float)
        gs = np.array([x.get("subset", "") for x in pool["items"]])
        sp = float(np.nanmean([spearman(cos[gs == g], gold[gs == g]) for g in dict.fromkeys(gs.tolist())]))
        res = {"pool_score": sp, "pool_metrics": {"cosine_spearman": sp}}
    elif t == "Clustering":
        X = enc([x["text"] for x in pool["items"]], None)
        nl = pool["n_levels"]
        cl = kmeans_runs(X, pool["items"], nl)
        agree = np.zeros(len(pool["items"]))
        cnt = np.zeros(len(pool["items"]))
        for gi in _group_index(pool["items"]):
            its = [pool["items"][i] for i in gi]
            for per in cl:
                for lv in range(nl):
                    mask, lab = level_labels(its, lv)
                    ids = per[lv][gi][mask]
                    maj = {}
                    for c in np.unique(ids):
                        maj[c] = np.bincount(lab[ids == c]).argmax()
                    hit = np.array([maj[c] == g for c, g in zip(ids, lab)], dtype=float)
                    agree[gi[mask]] += hit
                    cnt[gi[mask]] += 1
        agree /= np.maximum(cnt, 1)
        items = [{"id": x["id"], "agree": float(a), "clusters": [[int(per[lv][i]) for lv in range(nl)] for per in cl]}
                 for i, (x, a) in enumerate(zip(pool["items"], agree))]
        vm = cluster_score(X, pool["items"], nl)
        metrics = {"v_measure_kmeans": vm}
        if pool.get("mteb_ids"):          # mteb's own protocol on its downsampled documents (if all in pool)
            metrics["v_measure_mteb_protocol"] = mteb_cluster_protocol(X, pool)
        res = {"pool_score": vm, "pool_metrics": metrics}
        emb_out = X.astype(np.float16)
    elif t == "Summarization":
        mt = [s for x in pool["items"] for s in x["machine"]]
        ht = [s for x in pool["items"] for s in x["human"]]
        M, H = enc(mt, None), enc(ht, None)
        items, mi, hi = [], 0, 0
        for x in pool["items"]:
            m, h = M[mi:mi + len(x["machine"])], H[hi:hi + len(x["human"])]
            mi += len(x["machine"])
            hi += len(x["human"])
            pred = sim_matrix(m, h, "cosine").max(1)
            sp = spearman(x["relevance"], pred) if len(set(x["relevance"])) > 1 else float("nan")
            items.append({"id": x["id"], "spearman": sp, "pred": pred.tolist()})
        sc = float(np.nanmean([x["spearman"] for x in items]))
        res = {"pool_score": sc, "pool_metrics": {"cosine_spearman": sc}}
    elif t == "BitextMining":
        S1 = enc([x["sentence1"] for x in pool["items"]], None)
        keys = list(pool["bitext_texts"])
        T = enc([pool["bitext_texts"][k] for k in keys], None)
        kpos = {k: i for i, k in enumerate(keys)}
        cand_by = {sub: [kpos[k] for k in v] for sub, v in pool["bitext_candidates"].items()}
        items = []
        for i, x in enumerate(pool["items"]):
            tgt = kpos[x["target"]]
            cand = sorted(set(cand_by[x["subset"]]) | {tgt})
            sc = sim_matrix(S1[i:i + 1], T[cand], sim)[0]
            gold = sc[cand.index(tgt)]
            rank = int((sc > gold).sum()) + 1
            items.append({"id": x["id"], "correct": int(rank == 1), "rank": rank, "n_cand": len(cand)})
        g = [x["subset"] for x in pool["items"]]
        acc = group_mean([x["correct"] for x in items], g)
        res = {"pool_score": acc, "pool_metrics": {"accuracy": acc,
                                                    "mrr": group_mean([1 / x["rank"] for x in items], g)}}
    else:
        raise ValueError(t)
    res.update({"task": name, "type": t, "main_score": pool["main_score"], "n_items": len(items),
                "similarity": sim, "item_key": item_key, "items": items})
    return res, emb_out


def best_accuracy_threshold(s: np.ndarray, y: np.ndarray) -> tuple[float, float]:
    order = np.argsort(-s)
    ss, yy = s[order], y[order]
    tp = np.cumsum(yy)
    tn = (len(yy) - yy.sum()) - (np.arange(1, len(yy) + 1) - tp)
    acc = (tp + tn) / len(yy)
    j = int(np.argmax(acc))
    thr = float(ss[j]) if j + 1 >= len(ss) else float((ss[j] + ss[j + 1]) / 2)
    base = max(yy.mean(), 1 - yy.mean())
    if acc[j] < base:          # predicting all-negative is better
        return float(ss[0] + 1), float(base)
    return thr, float(acc[j])


def mteb_cluster_protocol(X: np.ndarray, pool: dict) -> Optional[float]:
    """mteb's ClusteringFast protocol (bootstrapped MiniBatchKMeans V-measure) on its own downsample."""
    try:
        from mteb.abstasks.AbsTaskClusteringFast import evaluate_clustering_bootstrapped
    except Exception:
        return None
    pos = {x["id"]: i for i, x in enumerate(pool["items"])}
    ids = [pos[i] for i in pool["mteb_ids"] if i in pos]
    if len(ids) < len(pool["mteb_ids"]) or len(ids) < 4:
        return None
    rng = random.Random(42)
    v = evaluate_clustering_bootstrapped(X[ids], [pool["items"][i]["labels"] for i in ids], n_clusters=10,
                                         cluster_size=pool.get("mteb_cluster_size", 16384), kmean_batch_size=512,
                                         max_depth=None, rng_state=rng)
    vals = [x for lv in v.values() for x in lv]
    return float(np.mean(vals))


# ----------------------------------------------------------------------------- subset scores (selection / eval)
class TaskScores:
    """Per-item scores of one task for M models, aligned on the pool's item ids."""

    def __init__(self, task: str, results: list[dict], embs: Optional[list[Optional[np.ndarray]]] = None,
                 pool: Optional[dict] = None):
        self.task, self.type = task, results[0]["type"]
        ids = [x["id"] for x in results[0]["items"]]
        for r in results[1:]:
            have = {x["id"] for x in r["items"]}
            ids = [i for i in ids if i in have]
        self.ids = ids
        self.M, self.N = len(results), len(ids)
        key = results[0].get("item_key") or ITEM_KEY[self.type]
        self.Y = np.zeros((self.M, self.N))
        for m, r in enumerate(results):
            byid = {x["id"]: x for x in r["items"]}
            for j, i in enumerate(ids):
                v = byid[i].get(key)
                self.Y[m, j] = np.nan if v is None else v
        self.pool_scores = np.array([r["pool_score"] for r in results], dtype=float)
        g = item_groups(pool) if pool is not None else {}
        self.group = np.array([g.get(i, "") for i in ids])
        self.multi = len(set(self.group.tolist())) > 1
        if self.type == "STS":
            b0 = {x["id"]: x for x in results[0]["items"]}
            self.gold = np.array([b0[i]["gold"] for i in ids], dtype=float)
        if self.type == "PairClassification":
            self.cos = np.zeros((self.M, self.N))
            for m, r in enumerate(results):
                byid = {x["id"]: x for x in r["items"]}
                self.cos[m] = [byid[i]["cos"] for i in ids]
            b0 = {x["id"]: x for x in pool["items"]} if pool else None
            self.label = np.array([b0[i]["label"] for i in ids]) if b0 else None
            self.w = pair_weights(pool, self.label) if b0 else None
        if self.type == "Clustering":
            self.clusters = []           # [m] -> array N x runs x levels
            for r in results:
                byid = {x["id"]: x for x in r["items"]}
                self.clusters.append(np.array([byid[i]["clusters"] for i in ids]))
            pitems = {x["id"]: x for x in pool["items"]}
            self.items = [pitems[i] for i in ids]
            self.n_levels = pool["n_levels"]
            self.labels = []
            for lv in range(self.n_levels):
                vocab = {}
                self.labels.append(np.array([vocab.setdefault(x["labels"][lv], len(vocab)) if len(x["labels"]) > lv
                                             else -1 for x in self.items]))
            self.embs = None
            if embs is not None and all(e is not None for e in embs):
                rows = {x["id"]: k for k, x in enumerate(pool["items"])}
                sel = [rows[i] for i in ids]
                self.embs = [e[sel].astype(np.float32) for e in embs]

    def _groups(self, cols: np.ndarray) -> list[np.ndarray]:
        if not self.multi:
            return [cols]
        gs = self.group[cols]
        return [cols[gs == g] for g in dict.fromkeys(gs.tolist())]

    def _sts(self, cols: np.ndarray, ms: np.ndarray) -> np.ndarray:
        from scipy.stats import rankdata
        g = rankdata(self.gold[cols])
        c = rankdata(self.Y[np.ix_(ms, cols)], axis=1)
        g = (g - g.mean()) / (g.std() + 1e-12)
        c = (c - c.mean(1, keepdims=True)) / (c.std(1, keepdims=True) + 1e-12)
        return (c * g).mean(1)

    def subset(self, cols: np.ndarray, models: Optional[np.ndarray] = None, exact: bool = False) -> np.ndarray:
        """Task score of an item subset for every model (cols: indices into self.ids); multi-subset tasks:
        mean over the hf_subsets present (set metrics: subsets with too few items are skipped, falling back
        to the pooled metric if none qualifies).
        exact=True: clustering re-clusters the subset (what the lite eval does); else the fast proxy
        (V-measure of the pool clusterings restricted to the subset)."""
        ms = np.arange(self.M) if models is None else np.asarray(models)
        cols = np.asarray(cols, dtype=int)
        t = self.type
        if t == "STS":
            if not self.multi:
                return self._sts(cols, ms)
            per = [self._sts(gc, ms) for gc in self._groups(cols) if len(gc) >= 3]
            return np.mean(per, 0) if per else self._sts(cols, ms)
        if t == "PairClassification":
            def ap(cc):
                w = None if self.w is None else self.w[cc]
                return np.array([average_precision(self.label[cc], self.cos[m, cc], w) for m in ms])
            if not self.multi:
                return ap(cols)
            per = [ap(gc) for gc in self._groups(cols)]
            per = [p for p in per if not np.isnan(p).any()]
            return np.mean(per, 0) if per else ap(cols)
        if t == "Clustering":
            if exact and self.embs is not None:
                its = [self.items[j] for j in cols]
                return np.array([cluster_score(self.embs[m][cols], its, self.n_levels) for m in ms])
            out = []
            for m in ms:
                cl = self.clusters[m]
                per_group = []
                for gc in self._groups(cols):
                    vals = []
                    for r in range(cl.shape[1]):
                        for lv in range(self.n_levels):
                            lab = self.labels[lv][gc]
                            ok = lab >= 0
                            if ok.sum() >= 2:
                                vals.append(v_measure_fast(lab[ok], cl[gc, r, lv][ok]))
                    if vals:
                        per_group.append(np.nanmean(vals))
                out.append(np.nanmean(per_group) if per_group else np.nan)
            return np.array(out)
        if not self.multi:
            return np.nanmean(self.Y[np.ix_(ms, cols)], axis=1)
        with np.errstate(all="ignore"):
            per = [np.nanmean(self.Y[np.ix_(ms, gc)], axis=1) for gc in self._groups(cols)]
        return np.nanmean(per, 0)


def load_scores(name: str, task: str) -> tuple[Optional[dict], Optional[np.ndarray]]:
    p = SCORES / name / f"{task}.json"
    if not p.exists():
        return None, None
    e = SCORES / name / f"{task}.emb.npy"
    return json.loads(p.read_text()), (np.load(e) if e.exists() else None)