Benchmark Almanac

KorSET

gemma2-9B-it 19.4% tops its leaderboard EXAONE-3.5-7.8B-Instruct 19.9% leads the models we track
Not in model cards

Models respond to Korean prompts adapted to local language, law, and culture. A model judge checks whether each response violates the relevant safety policy.

Selectstar, which ran three annual safety challenges, created KorSET. Its public card lists 9,249 prompts. The authors gate downloads and report five target models.

Kind
Deception and alignment
Level
Everyday
Scored
Graded by a model
Answers
Open-ended
Tasks
All public
Measures
Propensity
Behavior
Harmful use, Bias
Language
Korean
In and out
Text → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
In views
Content safety
Strongest model tested
Llama 3.1-8B #203 on the capability index
Last model added
March 2025
Scores refreshed
November 20th, 2026
5
Entries in its paper, 4 models we track

gemma2-9B-it tops the leaderboard with 19.4% EXAONE-3.5-7.8B-Instruct leads with 19.9% among models released by March 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline is the plain mean of the five risk-category rates in Table 3. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is EXAONE-3.5-7.8B-Instruct, at 19.9%. This view shows only models we track, each at its best.

The newest model with a score here came out in March 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 gemma2-9B-it Direct 19.4%
  2. 2 EXAONE-3.5-7.8B-Instruct GCG 19.9%
  3. 3 Gemma 3 12B GCG 20%
  4. 4 Qwen2.5-7B GCG 22.3%
  5. 5 Llama 3.1-8B GPTFuzzer 42.3%

Source: The KorSET paper

Source: The KorSET paper

#EntryScoreSettingDateSource
1 gemma2-9B-it 19.4% Direct Official leaderboard
2 EXAONE-3.5-7.8B-Instruct 19.9% GCG Dec 2024 Official leaderboard
3 Gemma 3 12B 20% GCG Mar 2025 Official leaderboard
4 Qwen2.5-7B 22.3% GCG Sep 2024 Official leaderboard
5 Llama 3.1-8B 42.3% GPTFuzzer Jul 2024 Official leaderboard
#ModelScoreSettingDateSource
1 EXAONE-3.5-7.8B-Instruct 19.9% GCG Dec 2024 Official leaderboard
2 Gemma 3 12B 20% GCG Mar 2025 Official leaderboard
3 Qwen2.5-7B 22.3% GCG Sep 2024 Official leaderboard
4 Llama 3.1-8B 42.3% GPTFuzzer Jul 2024 Official leaderboard

Rankings by part

We read 5 parts from the benchmark’s paper, each ranking models on one part of the test, such as Toxic Language, Unfair Representation and Misinformation Harms. EXAONE-3.5-7.8B-Instruct leads 2 of them, and 2 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Toxic Language

4 models we track have a score here.

  1. 1 Qwen2.5-7B ↑2 7.65%
  2. 2 EXAONE-3.5-7.8B-Instruct ↓1 7.73%
  3. 3 Gemma 3 12B ↓1 13.5%
  4. 4 Llama 3.1-8B 29.5%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Llama 3.1-8B
GPTFuzzer
44.8%
Qwen2.5-7B
GCG TAP AutoDAN GPTFuzzer
43.1%
Gemma 3 12B
GCG TAP AutoDAN GPTFuzzer
44%
EXAONE-3.5-7.8B-Instruct
GCG TAP AutoDAN GPTFuzzer
45.2%

Source: The KorSET paper

ModelEffortHarnessScoreCost per taskOutput tokens
EXAONE-3.5-7.8B-Instruct GPTFuzzer 45.2%
Llama 3.1-8B GCG 44.8%
Gemma 3 12B GPTFuzzer 44%
Llama 3.1-8B Direct 43.6%
Llama 3.1-8B AutoDAN 43.3%
Qwen2.5-7B GPTFuzzer 43.1%
Llama 3.1-8B TAP 42.6%
Llama 3.1-8B GPTFuzzer 42.3%
Gemma 3 12B AutoDAN 36.3%
EXAONE-3.5-7.8B-Instruct AutoDAN 35.1%
Qwen2.5-7B AutoDAN 33.6%
EXAONE-3.5-7.8B-Instruct TAP 32.2%
Qwen2.5-7B TAP 27%
Qwen2.5-7B Direct 24.4%
Gemma 3 12B TAP 24.1%
EXAONE-3.5-7.8B-Instruct Direct 22.6%
Qwen2.5-7B GCG 22.3%
Gemma 3 12B Direct 21.4%
Gemma 3 12B GCG 20%
EXAONE-3.5-7.8B-Instruct GCG 19.9%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Llama 3.1-8B
GPTFuzzer
44.8%
Qwen2.5-7B
GCG TAP AutoDAN GPTFuzzer
43.1%
Gemma 3 12B
GCG TAP AutoDAN GPTFuzzer
44%
EXAONE-3.5-7.8B-Instruct
GCG TAP AutoDAN GPTFuzzer
45.2%

Source: The KorSET paper

ModelEffortHarnessScoreCost per taskOutput tokens
EXAONE-3.5-7.8B-Instruct GPTFuzzer 45.2%
Llama 3.1-8B GCG 44.8%
Gemma 3 12B GPTFuzzer 44%
Llama 3.1-8B Direct 43.6%
Llama 3.1-8B AutoDAN 43.3%
Qwen2.5-7B GPTFuzzer 43.1%
Llama 3.1-8B TAP 42.6%
Llama 3.1-8B GPTFuzzer 42.3%
Gemma 3 12B AutoDAN 36.3%
EXAONE-3.5-7.8B-Instruct AutoDAN 35.1%
Qwen2.5-7B AutoDAN 33.6%
EXAONE-3.5-7.8B-Instruct TAP 32.2%
Qwen2.5-7B TAP 27%
Qwen2.5-7B Direct 24.4%
Gemma 3 12B TAP 24.1%
EXAONE-3.5-7.8B-Instruct Direct 22.6%
Qwen2.5-7B GCG 22.3%
Gemma 3 12B Direct 21.4%
Gemma 3 12B GCG 20%
EXAONE-3.5-7.8B-Instruct GCG 19.9%

3 models have held number one in the benchmark’s paper since 2023

Lab claims count from the day the card came out, independent runs from the model’s release.

Meta Qwen LG AI Research
Benchmark paper Not tested since Mar 2025

Source: The KorSET paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 EXAONE-3.5-7.8B-Instruct Qwen2.5-7B Qwen2.5-7B Qwen2.5-7B Llama 3.1-8B 0% 20% 40% 60% 80% 100% 2025 EXAONE-3.5-7.8B-Instruct Qwen2.5-7B

Source: Official KorSET leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Safety and alignment

All 636 in Safety and alignment
  • FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
  • StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. Inkling 98.6%
  • Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. Gemini 4 Carbon 0%
  • Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior. Inkling 2 Mini 8.30%
  • Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions. Inkling 2 Small 4.70%
  • Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior. Inkling 2 Mini 9.80%

Data audit

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 20 months ago, in March 2025, for Gemma 3 12B. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the benchmark’s paper (5 entries).
  • Pass Models scored 4 models we track, plus 1 other entry on its leaderboard.
  • Warning Newest score The newest came 20 months ago, in March 2025, for Gemma 3 12B.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Llama 3.1-8B (#203).
  • Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 20 rows for models we track and 1 other entry at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check