evaluate.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Supra-μBench — evaluation suite for very small base language models (<100M parameters).

Usage:
    python evaluate.py --model_dir path/to/model
                       [--batch_size 16] [--device auto] [--dtype float32]
                       [--max_length 512] [--tasks 1a,4b] [--output results.json]
                       [--show_levels] [--trust_remote_code] [--list_tasks]

All tasks are scored purely via log-likelihood (no generation).
"""

import argparse
import importlib.util
import json
import math
import sys
import time
from collections import defaultdict
from pathlib import Path

import torch
import torch.nn.functional as F
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

from tasks import BENCHMARK_NAME, VERSION, DOMAINS, DOMAIN_WEIGHTS, build_tasks

DTYPES = {"float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16}


# ----------------------------------------------------------------------------
# Model wrapper
# ----------------------------------------------------------------------------
class _LogitsOut:
    def __init__(self, logits):
        self.logits = logits

class _CausalAdapter(torch.nn.Module):
    """Make a custom module look like a causal LM: forward(input_ids=...).logits."""

    def __init__(self, inner, context):
        super().__init__()
        self.inner = inner
        self.config = type("Cfg", (), {
            "max_position_embeddings": context,
            "n_positions": context,
            "n_ctx": context,
            "seq_length": context,
        })()

    def forward(self, input_ids=None, attention_mask=None, **kwargs):
        ids = input_ids if input_ids is not None else kwargs.get("ids")
        try:
            out = self.inner(input_ids=ids, attention_mask=attention_mask)
        except Exception:
            try:
                out = self.inner(ids)
            except Exception:
                out = self.inner(ids=ids)
        if torch.is_tensor(out):
            logits = out
        elif isinstance(out, dict):
            logits = out.get("logits", out.get("logit"))
            if not torch.is_tensor(logits):
                raise TypeError(f"forward dict has no logits/logit tensor, keys={list(out)}")
        else:
            logits = getattr(out, "logits", None)
            if not torch.is_tensor(logits):
                raise TypeError(f"unsupported forward return: {type(out)}")
        return _LogitsOut(logits)

def _snapshot(model_dir):
    path = Path(model_dir)
    if path.is_dir():
        return path
    from huggingface_hub import snapshot_download
    return Path(snapshot_download(model_dir))

def _import_py(repo, filename):
    path = Path(repo) / filename
    spec = importlib.util.spec_from_file_location(f"_supra_{path.stem}", path)
    mod = importlib.util.module_from_spec(spec)
    sys.modules[spec.name] = mod  # dataclasses looks the class up in sys.modules during class body
    sys.path.insert(0, str(repo))
    try:
        spec.loader.exec_module(mod)
    except Exception:
        sys.modules.pop(spec.name, None)
        raise
    finally:
        sys.path.pop(0)
    return mod

def _load_state(model, repo):
    names = ("model.safetensors", "pytorch_model.bin", "model.bin", "model.pt", "model.pth")
    weight = next((Path(repo) / n for n in names if (Path(repo) / n).is_file()), None)
    if weight is None:
        cands = [p for p in Path(repo).iterdir()
                 if p.suffix in {".safetensors", ".bin", ".pt", ".pth"}
                 and "optim" not in p.name and "sched" not in p.name]
        if len(cands) != 1:
            raise FileNotFoundError(f"No unique weight file in {repo}: {cands}")
        weight = cands[0]
    if weight.suffix == ".safetensors":
        from safetensors.torch import load_file
        state = load_file(str(weight))
    else:
        state = torch.load(weight, map_location="cpu")
        if isinstance(state, dict):
            for key in ("state_dict", "model", "module"):
                if key in state and isinstance(state[key], dict):
                    state = state[key]
                    break
    # Strip a wrapper prefix only when every key has it (may be nested).
    for prefix in ("module.", "_orig_mod.", "model."):
        while state and all(k.startswith(prefix) for k in state):
            state = {k[len(prefix):]: v for k, v in state.items()}
    missing, unexpected = model.load_state_dict(state, strict=False)
    if missing or unexpected:
        print(f"  weight load: missing={len(missing)} unexpected={len(unexpected)} (kept)")
    return model

def _load_tokenizer(repo, trust_remote_code):
    repo = Path(repo)
    try:
        return AutoTokenizer.from_pretrained(repo, trust_remote_code=trust_remote_code)
    except Exception as exc:
        tok_json = repo / "tokenizer.json"
        if tok_json.is_file():
            from tokenizers import Tokenizer
            from transformers import PreTrainedTokenizerFast
            return PreTrainedTokenizerFast(tokenizer_object=Tokenizer.from_file(str(tok_json)))
        spm = repo / "tokenizer.model"
        if spm.is_file():
            from transformers import LlamaTokenizer
            return LlamaTokenizer(vocab_file=str(spm), legacy=False)
        vocab, merges = repo / "vocab.json", repo / "merges.txt"
        if vocab.is_file() and merges.is_file():
            from transformers import GPT2TokenizerFast
            return GPT2TokenizerFast(vocab_file=str(vocab), merges_file=str(merges))
        raise SystemExit(
            f"No usable tokenizer in {repo} ({exc}). "
            "Need tokenizer.json, tokenizer.model, or vocab.json+merges.txt."
        ) from exc

def _load_custom_repo(model_dir, dtype, trust_remote_code):
    repo = _snapshot(model_dir)
    cfg = json.loads((repo / "config.json").read_text(encoding="utf-8"))
    if (repo / "model.py").is_file():
        mod = _import_py(repo, "model.py")
        if hasattr(mod, "WorkspaceConfig") and hasattr(mod, "RecurrentWorkspace"):
            fields = {f.name for f in mod.WorkspaceConfig.__dataclass_fields__.values()}
            kwargs = {k: v for k, v in cfg.items() if k in fields}
            inner = mod.RecurrentWorkspace(mod.WorkspaceConfig(**kwargs))
            context = int(getattr(inner.cfg, "context", cfg.get("context", 512)))
            _load_state(inner, repo)
            tok = _load_tokenizer(repo, trust_remote_code)
            eos = cfg.get("eos_id", cfg.get("eos_token_id"))
            if tok.eos_token_id is None and eos is not None:
                tok.eos_token_id = int(eos)
            if tok.pad_token_id is None and tok.eos_token_id is not None:
                tok.pad_token_id = tok.eos_token_id
            return tok, _CausalAdapter(inner, context)
    for path in sorted(repo.glob("modeling_*.py")):
        mod = _import_py(repo, path.name)
        for obj in vars(mod).values():
            if isinstance(obj, type) and hasattr(obj, "from_pretrained"):
                model = obj.from_pretrained(repo, torch_dtype=dtype, trust_remote_code=True)
                return _load_tokenizer(repo, trust_remote_code), model
    raise ValueError(
        f"{model_dir} is not a Transformers model (no model_type) and has no loadable "
        "model.py (WorkspaceConfig/RecurrentWorkspace) or modeling_*.py."
    )

def load_causal_lm(model_dir, dtype, trust_remote_code):
    try:
        AutoConfig.from_pretrained(model_dir, trust_remote_code=trust_remote_code)
        tok = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=trust_remote_code)
        model = AutoModelForCausalLM.from_pretrained(
            model_dir, torch_dtype=dtype, trust_remote_code=trust_remote_code
        )
        return tok, model
    except Exception as exc:
        text = str(exc)
        custom = "model_type" in text or "Unrecognized model" in text or "trust_remote_code" in text
        if not trust_remote_code or not custom:
            raise
        print(f"  standard HF load failed ({exc.__class__.__name__}). Loading custom repo code ...")
        return _load_custom_repo(model_dir, dtype, trust_remote_code)

class LMScorer:
    def __init__(self, model_dir, device, dtype, max_length, batch_size, trust_remote_code=False):
        self.tokenizer, self.model = load_causal_lm(model_dir, dtype, trust_remote_code)
        try:
            self.model.to(device=device, dtype=dtype)
        except Exception as exc:
            print(f"  model.to(dtype=...) failed ({exc.__class__.__name__}); casting floating parameters only")
            self.model.to(device)
            with torch.no_grad():
                for p in self.model.parameters():
                    if p.is_floating_point() and p.dtype != dtype:
                        p.data = p.data.to(dtype=dtype)
        self.model.eval()
        self.device = device
        self.batch_size = batch_size

        # Every sequence starts with a prefix token (BOS, or EOS as fallback, like lm-eval).
        prefix = self.tokenizer.bos_token_id
        if prefix is None:
            prefix = self.tokenizer.eos_token_id
        if prefix is None:
            raise ValueError("Tokenizer has neither a BOS nor an EOS token; cannot build a start prefix.")
        self.prefix_id = prefix

        model_max = None
        for attr in ("max_position_embeddings", "n_positions", "n_ctx", "seq_length"):
            v = getattr(self.model.config, attr, None)
            if isinstance(v, int) and v > 0:
                model_max = v
                break
        self.max_length = min(max_length, model_max) if model_max else max_length
        self.n_params = sum(p.numel() for p in self.model.parameters())
        self._cache = {}

    # -- tokenization ---------------------------------------------------------
    def _encode(self, text):
        if not text:
            return []
        return self.tokenizer.encode(text, add_special_tokens=False)

    @staticmethod
    def _split_ws(context, continuation):
        """Move trailing whitespace of the context to the continuation (lm-eval convention)."""
        n = len(context) - len(context.rstrip())
        if n > 0:
            continuation = context[-n:] + continuation
            context = context[:-n]
        return context, continuation

    def _prepare(self, context, continuation):
        context, continuation = self._split_ws(context, continuation)
        if context:
            whole = self._encode(context + continuation)
            ctx = self._encode(context)
            if whole[: len(ctx)] == ctx and len(whole) > len(ctx):
                cont = whole[len(ctx):]
            else:  # tokenizer merged across the boundary -> fallback
                cont = self._encode(continuation)
        else:
            ctx, cont = [], self._encode(continuation)
        if not cont:
            raise ValueError(f"Empty continuation for context {context!r}")
        full = [self.prefix_id] + ctx + cont
        if len(full) - 1 > self.max_length:
            if len(cont) >= self.max_length:
                raise ValueError("Continuation is longer than max_length.")
            full = full[-(self.max_length + 1):]  # left-truncate the context
        return full, len(cont)

    # -- scoring ---------------------------------------------------------------
    @torch.inference_mode()
    def loglikelihood(self, requests):
        """requests: list of (context, continuation) -> list of summed log-probs (nats)."""
        unique = {}
        for i, req in enumerate(requests):
            unique.setdefault(req, []).append(i)
        keys = [k for k in unique if k not in self._cache]
        prepared = [self._prepare(c, x) for c, x in keys]
        order = sorted(range(len(keys)), key=lambda j: -len(prepared[j][0]))

        for b in range(0, len(order), self.batch_size):
            idx = order[b: b + self.batch_size]
            inputs = [prepared[j][0][:-1] for j in idx]
            T = max(len(x) for x in inputs)
            input_ids = torch.full((len(idx), T), self.prefix_id, dtype=torch.long)
            attn = torch.zeros((len(idx), T), dtype=torch.long)
            for r, x in enumerate(inputs):
                input_ids[r, : len(x)] = torch.tensor(x, dtype=torch.long)
                attn[r, : len(x)] = 1
            def _store(row_logits, j):
                full, n_cont = prepared[j]
                L = len(full) - 1
                lp = F.log_softmax(row_logits[L - n_cont: L].float(), dim=-1)
                tgt = torch.tensor(full[-n_cont:], device=lp.device).unsqueeze(1)
                self._cache[keys[j]] = lp.gather(1, tgt).sum().item()

            try:
                logits = self.model(input_ids=input_ids.to(self.device),
                                    attention_mask=attn.to(self.device)).logits
                for r, j in enumerate(idx):
                    _store(logits[r], j)
            except Exception as exc:
                print(f"  batch forward failed ({exc.__class__.__name__}); retrying one sequence at a time")
                for r, j in enumerate(idx):
                    one = torch.tensor([inputs[r]], device=self.device)
                    _store(self.model(input_ids=one).logits[0], j)

        results = [None] * len(requests)
        for req, idxs in unique.items():
            for i in idxs:
                results[i] = self._cache[req]
        return results

    @torch.inference_mode()
    def rolling_loglikelihood(self, text):
        """Sliding-window log-likelihood of a full text (nats), stride = max_length / 2."""
        seq = [self.prefix_id] + self._encode(text)
        L = self.max_length
        stride = max(1, L // 2)
        total, s = 0.0, 1
        while s < len(seq):
            e = min(s + stride, len(seq))
            start = max(0, e - L)
            inp = torch.tensor([seq[start:e]], device=self.device)
            logits = self.model(input_ids=inp).logits[0].float()
            lp = F.log_softmax(logits[s - 1 - start: e - 1 - start], dim=-1)
            tgt = torch.tensor(seq[s:e], device=self.device).unsqueeze(1)
            total += lp.gather(1, tgt).sum().item()
            s = e
        return total


# ----------------------------------------------------------------------------
# Task evaluation
# ----------------------------------------------------------------------------
def _softmax_prob(scores, label):
    m = max(scores)
    exps = [math.exp(s - m) for s in scores]
    return exps[label] / sum(exps)


def aggregate(records, metric):
    if metric == "group_acc":
        groups = defaultdict(list)
        for i, r in enumerate(records):
            # items without a group are scored as single units
            groups[r["group"] if r["group"] is not None else f"__single_{i}"].append(r)
        units = [{
            "correct": all(x["correct"] for x in g),
            "chance": math.prod(x["chance"] for x in g),
            "p_correct": math.prod(x["p_correct"] for x in g),
            "rr": sum(x["rr"] for x in g) / len(g),
            "level": max(x["level"] for x in g),
        } for g in groups.values()]
    else:
        units = records

    n = len(units)
    acc = sum(u["correct"] for u in units) / n
    chance = sum(u["chance"] for u in units) / n
    score = max(0.0, (acc - chance) / (1.0 - chance)) * 100.0
    se_acc = math.sqrt(acc * (1.0 - acc) / n)
    se_score = se_acc / (1.0 - chance) * 100.0

    by_lvl = defaultdict(list)
    for u in units:
        by_lvl[u["level"]].append(u)
    levels = {}
    for lvl in sorted(by_lvl):
        us = by_lvl[lvl]
        levels[lvl] = {
            "n": len(us),
            "acc": sum(u["correct"] for u in us) / len(us),
            "chance": sum(u["chance"] for u in us) / len(us),
        }
    return {
        "kind": "choice", "n_items": len(records), "n_units": n,
        "item_acc": sum(r["correct"] for r in records) / len(records),
        "acc": acc, "chance": chance, "score": score, "se": se_score,
        "p_correct": sum(u["p_correct"] for u in units) / n,
        "mrr": sum(u["rr"] for u in units) / n,
        "levels": levels,
    }


def evaluate_choice_task(scorer, task, dump=None):
    scoring = task["scoring"]
    default_null = task.get("null_context", "Answer:")
    requests, plan = [], []

    for it in task["items"]:
        entries = []
        if "contexts" in it:  # multi-context (partial scoring, Winograd style)
            nb = len(it["continuation"].encode("utf-8"))
            for ctx in it["contexts"]:
                requests.append((ctx, it["continuation"]))
                entries.append((len(requests) - 1, None, nb))
        else:
            for ch in it["choices"]:
                requests.append((it.get("context", ""), ch))
                ci, ni = len(requests) - 1, None
                if scoring == "pmi":
                    requests.append((it.get("null_context", default_null), ch))
                    ni = len(requests) - 1
                entries.append((ci, ni, len(ch.encode("utf-8"))))
        plan.append(entries)

    lls = scorer.loglikelihood(requests)

    records = []
    for it, entries in zip(task["items"], plan):
        scores = []
        for ci, ni, nb in entries:
            ll = lls[ci]
            if scoring == "sum":
                scores.append(ll)
            elif scoring == "mean_byte":
                scores.append(ll / max(nb, 1))
            elif scoring == "pmi":
                scores.append(ll - lls[ni])
            else:
                raise ValueError(f"Unknown scoring: {scoring}")
        label = it["label"]
        pred = max(range(len(scores)), key=lambda k: scores[k])
        rank = 1 + sum(1 for s in scores if s > scores[label])
        # mean_byte scores are per byte -> rescale to sequence level, otherwise P(corr) is ~uniform
        scale = sum(nb for _, _, nb in entries) / len(entries) if scoring == "mean_byte" else 1.0
        records.append({
            "level": it["level"],
            "correct": pred == label,
            "chance": 1.0 / len(scores),
            "p_correct": _softmax_prob([s * scale for s in scores], label),
            "rr": 1.0 / rank,
            "group": it.get("group"),
        })
        if dump is not None:
            dump.append({
                "task": task["name"], "level": it["level"], "group": it.get("group"),
                "context": it.get("context", it.get("contexts")),
                "choices": it.get("choices", it.get("continuation")),
                "scores": [round(s, 4) for s in scores], "label": label, "pred": pred,
                "correct": pred == label,
            })
    return aggregate(records, task.get("metric", "acc"))


def evaluate_bpb_task(scorer, task):
    total_ll, total_bytes, per_text = 0.0, 0, []
    for text in task["texts"]:
        ll = scorer.rolling_loglikelihood(text)
        nb = len(text.encode("utf-8"))
        total_ll += ll
        total_bytes += nb
        per_text.append(-ll / (math.log(2) * nb))
    bpb = -total_ll / (math.log(2) * total_bytes)
    lo, hi = task["anchors"]["floor"], task["anchors"]["ceiling"]
    score = 100.0 * (math.log(lo) - math.log(bpb)) / (math.log(lo) - math.log(hi))
    return {"kind": "bpb", "n_items": len(task["texts"]), "bpb": bpb,
            "bpb_per_text": per_text, "score": min(100.0, max(0.0, score)), "se": 0.0}


# ----------------------------------------------------------------------------
# Index
# ----------------------------------------------------------------------------
def compute_index(tasks, results):
    by_dom = defaultdict(list)
    for t in tasks:
        by_dom[t["domain"]].append(results[t["name"]])
    dom_scores, dom_se = {}, {}
    for d, rs in by_dom.items():
        k = len(rs)
        dom_scores[d] = sum(r["score"] for r in rs) / k
        dom_se[d] = math.sqrt(sum(r["se"] ** 2 for r in rs)) / k
    wsum = sum(DOMAIN_WEIGHTS[d] for d in dom_scores)
    index = sum(DOMAIN_WEIGHTS[d] * s for d, s in dom_scores.items()) / wsum
    index_se = math.sqrt(sum((DOMAIN_WEIGHTS[d] / wsum) ** 2 * dom_se[d] ** 2 for d in dom_scores))
    return dom_scores, dom_se, index, index_se


# ----------------------------------------------------------------------------
# Reporting
# ----------------------------------------------------------------------------
def print_report(meta, tasks, results, dom_scores, dom_se, index, index_se, show_levels):
    line = "=" * 96
    print("\n" + line)
    print(f"{BENCHMARK_NAME} v{VERSION}")
    print(f"Model: {meta['model_dir']}  |  Params: {meta['n_params'] / 1e6:.1f}M  |  "
          f"max_length: {meta['max_length']}  |  device: {meta['device']}  |  {meta['runtime_s']:.1f}s")
    print(line)
    print(f"{'Task':<32}{'Dom':<5}{'N':>5}{'Acc%':>8}{'Chance%':>9}{'NormAcc':>9}{'±SE':>7}"
          f"{'P(corr)':>9}{'MRR':>7}")
    print("-" * 96)
    for t in tasks:
        r = results[t["name"]]
        if r["kind"] == "bpb":
            print(f"{t['name']:<32}{t['domain']:<5}{r['n_items']:>5}   BPB = {r['bpb']:.4f}"
                  f"{'':>6}{r['score']:>9.1f}   (anchor-normalized)")
            continue
        print(f"{t['name']:<32}{t['domain']:<5}{r['n_units']:>5}{r['acc'] * 100:>8.1f}"
              f"{r['chance'] * 100:>9.1f}{r['score']:>9.1f}{r['se']:>7.1f}"
              f"{r['p_correct']:>9.3f}{r['mrr']:>7.3f}")
        if show_levels:
            lv = "  ".join(f"L{l}: {v['acc'] * 100:.0f}% (n={v['n']}, c={v['chance'] * 100:.0f}%)"
                           for l, v in r["levels"].items())
            print(f"{'':<6}{lv}")
    print("-" * 96)
    print("Domain scores (0-100, chance-normalized):")
    for d in DOMAINS:
        if d in dom_scores:
            print(f"  {d} {DOMAINS[d]:<34} w={DOMAIN_WEIGHTS[d]:.2f}   "
                  f"{dom_scores[d]:6.1f}  ± {dom_se[d]:.1f}")
    print(line)
    print(f"  μBench-Intelligence-Index:  {index:.2f}  ± {index_se:.2f}  (SE, item sampling)")
    print(line)
    print("Note: treat index differences < 2*sqrt(SE_a^2 + SE_b^2) as noise.\n")


# ----------------------------------------------------------------------------
# Main
# ----------------------------------------------------------------------------
def resolve_device(arg):
    if arg != "auto":
        return torch.device(arg)
    if torch.cuda.is_available():
        return torch.device("cuda")
    if getattr(torch.backends, "mps", None) is not None and torch.backends.mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")


def parse_args():
    p = argparse.ArgumentParser(description=f"{BENCHMARK_NAME} evaluation suite")
    p.add_argument("--model_dir", help="Local model folder or Hugging Face model id (e.g. org/name); must include the tokenizer")
    p.add_argument("--batch_size", type=int, default=16)
    p.add_argument("--device", default="auto", help="auto | cpu | cuda | cuda:0 | mps")
    p.add_argument("--dtype", default="float32", choices=list(DTYPES))
    p.add_argument("--max_length", type=int, default=512)
    p.add_argument("--tasks", default=None, help="Comma-separated task names or prefixes, e.g. '1a,4b'")
    p.add_argument("--output", default=None, help="Write results as JSON to this path")
    p.add_argument("--dump", default=None, help="Write per-item predictions as JSONL to this path")
    p.add_argument("--show_levels", action="store_true", help="Print accuracy per difficulty level")
    p.add_argument("--trust_remote_code", action="store_true",
                   help="Load custom modeling code (configuration_*.py, modeling_*.py) from a local folder or the HF Hub")
    p.add_argument("--list_tasks", action="store_true", help="List tasks and exit")
    return p.parse_args()


def main():
    args = parse_args()
    tasks = build_tasks()

    if args.list_tasks:
        for t in tasks:
            n = len(t.get("items", t.get("texts", [])))
            print(f"{t['name']:<32}{t['domain']:<5}{n:>4}  {t['description']}")
        return
    if not args.model_dir:
        raise SystemExit("--model_dir is required (local folder or HF model id).")

    if args.tasks:
        wanted = [w.strip() for w in args.tasks.split(",") if w.strip()]
        tasks = [t for t in tasks if t["name"] in wanted or t["name"].split("_")[0] in wanted]
        if not tasks:
            raise SystemExit(f"No tasks match {wanted}.")

    device = resolve_device(args.device)
    print(f"Loading model from {args.model_dir} on {device} ({args.dtype}) ...")
    scorer = LMScorer(args.model_dir, device, DTYPES[args.dtype], args.max_length,
                      args.batch_size, args.trust_remote_code)

    results, t0 = {}, time.time()
    dump_rows = [] if args.dump else None
    for t in tasks:
        ts = time.time()
        if t.get("kind") == "bpb":
            results[t["name"]] = evaluate_bpb_task(scorer, t)
        else:
            results[t["name"]] = evaluate_choice_task(scorer, t, dump_rows)
        print(f"  [{t['name']}] score={results[t['name']]['score']:.1f}  ({time.time() - ts:.1f}s)")

    dom_scores, dom_se, index, index_se = compute_index(tasks, results)
    meta = {
        "benchmark": BENCHMARK_NAME, "version": VERSION, "model_dir": args.model_dir,
        "n_params": scorer.n_params, "max_length": scorer.max_length, "device": str(device),
        "dtype": args.dtype, "runtime_s": time.time() - t0,
    }
    print_report(meta, tasks, results, dom_scores, dom_se, index, index_se, args.show_levels)

    if args.output:
        out = {
            "meta": meta,
            "tasks": {t["name"]: {"domain": t["domain"], "format": t["format"],
                                  "scoring": t.get("scoring"), **results[t["name"]]} for t in tasks},
            "domains": {d: {"name": DOMAINS[d], "weight": DOMAIN_WEIGHTS[d],
                            "score": dom_scores[d], "se": dom_se[d]} for d in dom_scores},
            "mubench_intelligence_index": index,
            "mubench_intelligence_index_se": index_se,
        }
        with open(args.output, "w", encoding="utf-8") as f:
            json.dump(out, f, indent=2, ensure_ascii=False)
        print(f"Results written to {args.output}")

    if args.dump:
        with open(args.dump, "w", encoding="utf-8") as f:
            for row in dump_rows:
                f.write(json.dumps(row, ensure_ascii=False) + "\n")
        print(f"Per-item predictions written to {args.dump}")


if __name__ == "__main__":
    main()