DaehanKim/korean-jev-benchmark

Korean JEV Benchmark

25,418 Korean text-only evaluation items from 9 public datasets and 22 subsets, normalized to finite-choice decisions such as A, B, or C.

Baseline results · 9 October 2026

All ten measured baseline configurations, including Vega and hosted JEV, appear in both the table and radar.

Farther outward is better on every radar axis. Throughput is normalized across these ten models; accuracy and order agreement are raw percentages. HTTP p50 uses a reversed linear millisecond axis: 3,000 ms at the center, 0 ms at the rim, with 500 ms ticks. Farther outward means lower latency.

Model	Precision / service	C*	Decisions/s ↑	HTTP p50 ms ↓	Korean accuracy ↑	English accuracy ↑	Order KO ↑	Order EN ↑	Order macro ↑
Gemma 4 31B	FP8	64	45.2	1343.3	94.40%	57.64%	97.68%	55.00%	76.34%
Qwen3.8 27B	FP8	64	52.9	1162.4	89.02%	60.40%	95.29%	61.38%	78.34%
Gemma 4 26B-A4B	FP8	128	200.4	614.6	90.37%	45.36%	95.63%	41.75%	68.69%
Gemma 4 12B	FP8	64	114.4	533.4	89.58%	39.69%	92.72%	36.46%	64.59%
Cygnet · Gemma 12B	BF16	64	49.9	1244.7	89.18%	54.84%	95.47%	58.73%	77.10%
Mica v0.1 4B	BF16	64	152.0	432.7	81.47%	52.96%	89.83%	61.25%	75.54%
K-Decision 4B	BF16	256	102.7	2451.4	85.41%	48.94%	93.44%	53.12%	73.28%
Qwen3.5 4B	BF16	64	185.6	332.5	78.77%	46.24%	77.49%	44.75%	61.12%
Decision 2.0 Vega 27B	BF16 / FP32 head · official max_speed	1	11.1	71.3	92.38%	64.59%	96.84%	73.78%	85.31%
JEV 1.13 · hosted API	Hosted API	64	88.4	543.7	88.93%	82.50%	96.39%	89.28%	92.83%

Local models use one RTX PRO 6000 Blackwell Server 96GB GPU, with 256 warmup requests per timed pass. JEV is hosted through OpenRouter; its hardware is undisclosed. C* maximizes observed throughput across concurrency 1/8/32/64/128/256/512 on the same 2,048-item workload. HTTP p50 is measured at that C* and includes server/API queueing and transport after submission; it excludes local caller-slot wait.

Accuracy uses 25,418 Korean items and 12,032 English MMLU-Pro items (zero-shot direct decisions, without reasoning). Order agreement compares original and exact-reverse choices after semantic alignment; Order macro gives the two languages equal weight. Agreement can include consistently wrong answers.

Measurement details · Complete baseline metrics.

Earlier results

4 October local-model and hosted JEV API results use queue-inclusive timing and remain archived as earlier measurements.

Load and evaluate

import json
from datasets import load_dataset

ds = load_dataset("DaehanKim/korean-jev-benchmark", split="test")
row = ds[0]
model_input = {
    "task": row["instruction"],
    "item": json.loads(row["item_json"]),
    "choices": row["choices"],
}
gold = row["answer_id"]  # Evaluation only; never include in the prompt.

Use the released choice order and score exact-match choice accuracy. There is one evaluation split, with 2–11 choices per item (supported scope: up to 26). There is no training or calibration split. Source and subset provenance are retained for breakdowns; keep answers, provenance and audit judgments out of model prompts.

About the data

Sources: KLUE, KoBEST, Belebele, FunctionChat-Bench, APEACH, Korean Hate Speech/BEEP, SIB-200, EnterpriseOps-KO Policy, and KD-Lawset-KO. Tasks cover inference, comprehension, commonsense, classification, policy decisions, tool selection and questions about supplied statutes. Evaluation compares choice IDs; it does not use a reward model or LLM judge.

Labels were audited with GPT-6 Luna and Codex text adjudication; original gold labels were preserved and unresolved cases excluded. This is a curated derivative of public datasets, not an independent human annotation study or a guaranteed uncontaminated test set.

Dataset sources, construction, full review process, schema and limitations → detail.md

Licenses remain source-dependent: CC BY-SA 4.0, CC BY 4.0 and Apache-2.0. No CC BY-NC 4.0 data is included. See LICENSE, source attributions, and the citation.