mbppplus/evaluator.py

#!/usr/bin/env python3
"""Zevo custom evaluator for MBPP+ val (378), pass@1.

Hub: Tommysha/zero-code-l1-val config mbppplus split validation.
Rows: id, prompt, entry_point, test_imports, test_list, test (EvalPlus plus harness).
Does not use the gold `code` column.

Usage: python3 evaluator.py PREDICTIONS_CSV SCORING_SET METRICS_JSON
Stdlib only (pyarrow only if SCORING_SET is parquet). Timeout 10 s per problem.
If numpy is missing, a tiny shim satisfies `import numpy as np` in the plus harness.
"""
import ast
import csv
import json
import multiprocessing
import os
import sys
import traceback
from concurrent.futures import ProcessPoolExecutor

TIMEOUT = 10


def extract_code(output: str) -> str:
    lines = (output or "").split("\n")
    idx = [i for i, line in enumerate(lines) if "```" in line]
    if len(idx) >= 2:
        body = "\n".join(lines[idx[-2] + 1: idx[-1]])
        if body.lstrip().startswith("python"):
            body = "\n".join(body.split("\n")[1:])
        return body
    return (output or "").strip()


def _numpy_shim():
    class _NP:
        ndarray = type("ndarray", (), {})
        float64 = "float64"
        float32 = "float32"

        @staticmethod
        def allclose(a, b, rtol=1e-07, atol=1e-06, equal_nan=False):
            def _flat(x):
                if isinstance(x, (list, tuple, set)):
                    out = []
                    for i in x:
                        out.extend(_flat(i))
                    return out
                return [x]
            aa, bb = _flat(a), _flat(b)
            if len(aa) != len(bb):
                return False
            for x, y in zip(aa, bb):
                try:
                    xf, yf = float(x), float(y)
                except (TypeError, ValueError):
                    if x != y:
                        return False
                    continue
                if xf != xf or yf != yf:
                    if not equal_nan:
                        return False
                    continue
                scale = max(abs(yf), abs(xf))
                if abs(xf - yf) > atol + rtol * scale:
                    return False
            return True

    return _NP()


def _run_one(code, test, test_list, test_imports, result):
    try:
        ns = {}
        for imp in test_imports or []:
            exec(imp, ns, ns)
        exec(code, ns, ns)
        plus = (test or "").strip()
        if plus:
            if "numpy" not in sys.modules:
                try:
                    import numpy  # noqa: F401
                except Exception:
                    sys.modules["numpy"] = _numpy_shim()
            exec(plus, ns, ns)
        else:
            for stmt in test_list or []:
                exec(stmt, ns, ns)
        result.append(True)
    except Exception:
        result.append(False)


def check(args):
    code, test, test_list, test_imports = args
    if not (code or "").strip():
        return False
    try:
        ast.parse(code)
    except SyntaxError:
        return False
    manager = multiprocessing.Manager()
    result = manager.list()
    p = multiprocessing.Process(
        target=_run_one, args=(code, test, test_list, test_imports, result)
    )
    p.start()
    p.join(timeout=TIMEOUT + 2)
    if p.is_alive():
        p.kill()
        p.join(1)
    return bool(result) and result[0] is True


def load_rows(path):
    ext = os.path.splitext(path)[1].lower()
    if ext == ".parquet":
        import pyarrow.parquet as pq
        return pq.read_table(path).to_pylist()
    if ext == ".jsonl":
        return [json.loads(l) for l in open(path, encoding="utf-8") if l.strip()]
    if ext == ".json":
        d = json.load(open(path, encoding="utf-8"))
        return d if isinstance(d, list) else d.get("data", d.get("rows"))
    csv.field_size_limit(sys.maxsize)
    return list(csv.DictReader(open(path, newline="", encoding="utf-8")))


def _list_field(v):
    if v is None:
        return []
    if isinstance(v, list):
        return v
    if isinstance(v, str):
        s = v.strip()
        if not s:
            return []
        try:
            parsed = json.loads(s)
            if isinstance(parsed, list):
                return parsed
        except Exception:
            pass
        return [s]
    return list(v)


def main():
    pred_path, set_path, out_path = sys.argv[1:4]
    csv.field_size_limit(sys.maxsize)
    preds = {r["id"]: r.get("prediction") or "" for r in csv.DictReader(open(pred_path, newline="", encoding="utf-8"))}
    rows = load_rows(set_path)
    jobs = []
    for r in rows:
        jobs.append((
            extract_code(preds.get(str(r["id"]), "")),
            r.get("test") or "",
            _list_field(r.get("test_list")),
            _list_field(r.get("test_imports")),
        ))
    workers = int(os.environ.get("LCB_EVAL_WORKERS", min(8, os.cpu_count() or 1)))
    with ProcessPoolExecutor(max_workers=workers) as ex:
        ok = list(ex.map(check, jobs))
    total = len(rows)
    correct = sum(ok)
    out = {
        "score": correct / total if total else 0.0,
        "correct": correct,
        "total": total,
        "missing_predictions": sum(str(r["id"]) not in preds for r in rows),
    }
    json.dump(out, open(out_path, "w", encoding="utf-8"))
    print(json.dumps(out))


if __name__ == "__main__":
    main()