K2S3-Llama2-13B-v1.0: benchmark results
K2S3 release published by Changgil: a full-parameter supervised fine-tune of meta-llama/Llama-2-13b-chat-hf using Korean dictionary, KULLM, thesis-abstract and AI Hub data, with added Korean tokenizer tokens. It is distinct from the separately released LoRA-trained test model. Provider: Changgil. Access: Open.
Aggregate Rating 1317 ± 34, rank #8398 of 10916 rated models, from 6 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - WinoGrande | 74.35 | Accuracy (%) (5-shot) | 36.1 |
| Open LLM Leaderboard v1 - HellaSwag | 78.57 | Normalized accuracy (%) (10-shot) | 34.5 |
| Open LLM Leaderboard v1 - MMLU | 50.99 | Accuracy (%) (5-shot) | 30.3 |
| Open LLM Leaderboard v1 - TruthfulQA MC2 | 45.17 | MC2 (%) (0-shot) | 29.9 |
| Open LLM Leaderboard v1 - ARC Challenge | 51.45 | Normalized accuracy (%) (25-shot) | 25.9 |
| Open LLM Leaderboard v1 - GSM8K | 4.55 | Accuracy (%) (5-shot) | 23 |
Interactive version: theaggregate.ai/model?slug=k2s3-llama2-13b-v1-0 · How It Works · Data refreshed daily, snapshot 2026-11-08.