작은 오염 점검을 통해 인공지능 코딩 벤치마크 배우기

OpenAI published “Separating signal from noise in coding evaluations” on July 8, 2026, describing reliability concerns in SWE-Bench Pro. The beginner lesson is bigger than one leaderboard: a score is only meaningful when the tasks, grading, and possible training overlap are understood.

You can learn the basic idea with a tiny local exercise. Suppose train.txt contains filenames or issue phrases seen during development, and eval.txt contains benchmark prompts.

from pathlib import Path
import re

def tokens(text):
    return set(re.findall(r"[a-z0-9_]{4,}", text.lower()))

train = tokens(Path("train.txt").read_text())
for line in Path("eval.txt").read_text().splitlines():
    words = tokens(line)
    overlap = words & train
    ratio = len(overlap) / max(1, len(words))
    print(f"{ratio:.2f}\t{sorted(overlap)[:6]}\t{line[:60]}")

Enter fullscreen mode Exit fullscreen mode

Try these fixtures:

# train.txt
fix websocket reconnect timer in stream_client
normalize windows path in cache loader

# eval.txt
Repair the reconnect timer in stream_client after a websocket closes.
Add keyboard navigation to the settings dialog.

Enter fullscreen mode Exit fullscreen mode

The first line should show much more lexical overlap. That does not prove contamination. It is only a flag for human review. Real investigations need dataset provenance, timestamps, deduplication methods, repository history, and an analysis of whether the overlap reveals the answer.

A four-part benchmark note

Whenever you quote a coding score, record:

• exact benchmark version and split;
• model and tool configuration;
• grading method and known exclusions;
• overlap checks and unresolved limitations.

I apply the same discipline when evaluating coding products. I use MonkeyCode, but I would not recommend it from a vendor score alone. I recommend trying a few versioned tasks from your own repositories and keeping expected tests beside the prompt. Its open-source option is useful when you want to inspect or self-host the surrounding workflow; the hosted SaaS is useful when you want to start without operating that stack.

That is a workflow recommendation, not a claim that I measured one model as universally best.

Disclosure: I’m a MonkeyCode user sharing my own experience, not affiliated with the project.

The main learning outcome is simple: leaderboards are inputs to an experiment. They are not substitutes for a test set that represents your code, constraints, and definition of a correct change.