File size: 6,173 Bytes
8b97eb8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 | """RealSR v3 scoring harness — two parallel scores (see README → How scores are defined).
numeric_score — deterministic, reference-relative. Computed by `evaluate_numeric.py`
(`score_one`), exposed here as `evaluate_on_test()`. Type II
is 3-seed averaged. Needs only the task's
data + reference_metrics.json — no API key, no network.
validity_score — produced SEPARATELY by a Claude-Code (cc) subagent that
executes the formula on the data and scores the task's
`validity_rubrics` (see VALIDITY_JUDGE.md). Not computed here.
"""
from __future__ import annotations
import importlib.util
import json
import os
import sys
import tempfile
from pathlib import Path
from typing import Any, Dict, Tuple
_PKG_DIR = Path(__file__).resolve().parent
if str(_PKG_DIR) not in sys.path:
sys.path.insert(0, str(_PKG_DIR))
import evaluate_numeric as _ev # noqa: E402 official harness scorer (numeric)
import eval_formula as _ef # noqa: E402 execution core (load_flat/clusters)
_DIAG_METRICS = ("rmse", "mae", "mse", "mdae", "smape", "mape", "log_mae", "r2")
__all__ = ["evaluate_on_test"]
def _load_submission_module(code: str) -> Tuple[Any, str]:
"""Write the submitted module text to a real temp .py and import it, then
inject the harness-contract defaults the LLM may have omitted. A real file
is used so source-reading tools work. Returns (module, temp_path); the
caller unlinks temp_path."""
fd, path = tempfile.mkstemp(suffix=".py", prefix="_llm_submission_")
with os.fdopen(fd, "w") as fh:
fh.write(code)
spec = importlib.util.spec_from_file_location(f"_llm_submission_{Path(path).stem}", path)
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)
if not hasattr(mod, "USED_INPUTS"):
mod.USED_INPUTS = []
if not hasattr(mod, "LAW_CONSTANTS"):
mod.LAW_CONSTANTS = {}
if not hasattr(mod, "OTHER_CONSTANTS"):
mod.OTHER_CONSTANTS = {}
if not hasattr(mod, "LOCAL_FITTABLE"):
mod.LOCAL_FITTABLE = {}
elif isinstance(mod.LOCAL_FITTABLE, (list, tuple, set)):
mod.LOCAL_FITTABLE = {k: {} for k in mod.LOCAL_FITTABLE}
return mod, path
def evaluate_on_test(submitted_code: str, task, **_ignored) -> Dict[str, Any]:
"""Score `submitted_code` against `task` — numeric_score only,
via the official numeric harness `score_one` (Type II 3-seed averaged). validity_score
is produced separately by the cc judge (VALIDITY_JUDGE.md).
Returns: status, contract_ok, numeric_score, numeric_score_std,
numeric_score_per_seed, metric, best_reference_id, raw_metric, metrics,
violations, error.
"""
task_dir = Path(task.task_dir)
meta = _ev.load_task(task_dir)
task_type = meta.get("type", "typeII")
ref_path = _ev.reference_metrics_path(task_dir) # private scoring/ tree
if not ref_path.exists():
return {"status": "no_reference", "contract_ok": None,
"error": f"{ref_path} missing — run `evaluate_numeric.py reference`",
"numeric_score": 0.0, "raw_numeric_score": None,
"numeric_score_std": 0.0,
"numeric_score_per_seed": [0.0] if task_type == "typeI" else [0.0] * _ev.N_SEEDS}
ref_metrics = json.loads(ref_path.read_text())
try:
mod, tmp_path = _load_submission_module(submitted_code)
except Exception as e:
return {"status": "compile_error", "contract_ok": False,
"error": f"{type(e).__name__}: {e}",
"numeric_score": 0.0, "raw_numeric_score": None,
"numeric_score_std": 0.0,
"numeric_score_per_seed": [0.0] if task_type == "typeI" else [0.0] * _ev.N_SEEDS}
try:
# Context-ablation arms rename the columns the agent saw (x_1, ...). The
# harness reads the REAL test.csv, so map USED_INPUTS back to real names
# (order preserved → predict's positional X columns stay correct).
v2r = getattr(task, "view_to_real", None)
if v2r and getattr(mod, "USED_INPUTS", None):
mod.USED_INPUTS = [v2r.get(c, c) for c in mod.USED_INPUTS]
data = _ef.load_flat(task_dir) if task_type == "typeI" else _ef.load_clusters(task_dir)
# numeric_score needs only reference_metrics.json (anchors); the reference
# baseline .py files are not shipped, so the registry is empty.
r = _ev.score_one(mod, "submission", data, ref_metrics, meta, {})
finally:
try:
os.unlink(tmp_path)
except OSError:
pass
if not r.get("contract_ok"):
return {"status": "contract_fail", "contract_ok": False,
"violations": r.get("violations"),
"numeric_score": r.get("numeric_score", 0.0),
"raw_numeric_score": r.get("raw_numeric_score"),
"numeric_score_std": r.get("numeric_score_std", 0.0),
"numeric_score_per_seed": r.get("numeric_score_per_seed"),
"metric": _ev.task_metric(meta)}
failed = bool(r.get("failed")) if task_type == "typeI" \
else r.get("n_clusters_failed") == len(data["cluster_ids"])
score = r.get("score") or {}
metric = score.get("metric") or _ev.task_metric(meta)
raw_metric = r.get("raw_metric")
out: Dict[str, Any] = {
"status": r.get("status") or ("ok" if not failed else "exec_error"),
"contract_ok": True,
"error": r.get("error"),
"numeric_score": r.get("numeric_score"),
"raw_numeric_score": r.get("raw_numeric_score"),
"numeric_score_std": r.get("numeric_score_std"),
"numeric_score_per_seed": r.get("numeric_score_per_seed"),
"metric": metric,
"best_reference_id": score.get("best_reference_id"),
"raw_metric": raw_metric,
"metrics": {metric: raw_metric} if isinstance(raw_metric, (int, float)) else {},
}
if task_type != "typeI":
out["per_cluster_score"] = score.get("per_cluster_score")
out["n_clusters_scored"] = score.get("n_clusters_scored")
return out
|