Model Card for Model ID


Open Ko LLM Leaderboard Season 2 ๐Ÿ† Rank-1 2024/11/01~2024/12/28


AI ์™€ ๋น…๋ฐ์ดํ„ฐ ๋ถ„์„ ์ „๋ฌธ ๊ธฐ์—…์ธ Linkbricks์˜ ๋ฐ์ดํ„ฐ์‚ฌ์ด์–ธํ‹ฐ์ŠคํŠธ์ธ ์ง€์œค์„ฑ(Saxo) ๋ฐ•์‚ฌ๊ฐ€ gemma-2-27b-it ๋ฒ ์ด์Šค๋ชจ๋ธ์„ H100-80G 8๊ฐœ๋ฅผ ํ†ตํ•ด SFT->DPO ํŒŒ์ธ ํŠœ๋‹์„ ํ•œ ํ•œ๊ธ€ ์–ธ์–ด ๋ชจ๋ธ๋กœ ํ•œ๊ตญ์–ด-์ค‘๊ตญ์–ด-์˜์–ด-์ผ๋ณธ์–ด ๊ต์ฐจ ํ•™์Šต ๋ฐ์ดํ„ฐ์™€ ๋กœ์ง€์ปฌ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•˜์—ฌ ํ•œ์ค‘์ผ์˜ ์–ธ์–ด ๊ต์ฐจ ์ฆ๊ฐ• ์ฒ˜๋ฆฌ์™€ ๋ณต์žกํ•œ ํ•œ๊ธ€ ๋…ผ๋ฆฌ ๋ฌธ์ œ ์—ญ์‹œ ๋Œ€์‘ ๊ฐ€๋Šฅํ•˜๋„๋ก ํ›ˆ๋ จํ•œ ๋ชจ๋ธ์ด๋ฉฐ ํ† ํฌ๋‚˜์ด์ €๋Š” ๋‹จ์–ด ํ™•์žฅ ์—†์ด ๋ฒ ์ด์Šค ๋ชจ๋ธ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉ. ํŠนํžˆ ๊ณ ๊ฐ ๋ฆฌ๋ทฐ๋‚˜ ์†Œ์…œ ํฌ์ŠคํŒ… ๊ณ ์ฐจ์› ๋ถ„์„ ๋ฐ ์ฝ”๋”ฉ๋“ฑ์ด ๊ฐ•ํ™”๋œ ๋ชจ๋ธ
-Deepspeed Stage=3, rslora ๋ฐ BAdam Layer Mode ์‚ฌ์šฉ
-ollama run benedict/linkbricks-gemma2-korean:27b

Benchmark (Open Ko LLM Leader Board Season 2 : No. 1)
Model : Saxo/Linkbricks-Horizon-AI-Korean-Gemma-2-sft-dpo-27B
Average : 51.37
Ko-GPQA : 25.25
Ko-Winogrande : 68.27
Ko-GSM8k : 70.96
Ko-EQ Bench : 50.25
Ko-IFEval : 49.84
KorNAT-CKA : 34.59
KorNAT-SVA : 48.42
Ko-Harmlessness : 65.66
Ko-Helpfulness : 49.12




Dr. Yunsung Ji (Saxo), a data scientist at Linkbricks, a company specializing in AI and big data analytics, fine-tuned the gemma-2-27b-it base model with SFT->DPO using four H100-80Gs. It is a Korean language model trained to handle complex Korean logic problems through Korean-Chinese-English-Japanese cross-training data and logical data, and Tokenizer uses the base model without word expansion.

www.linkbricks.com, www.linkbricks.vc

Downloads last month
2,580
Safetensors
Model size
27.2B params
Tensor type
BF16
ยท
Inference Providers NEW
This model is not currently available via any of the supported third-party Inference Providers, and the model is not deployed on the HF Inference API.

Model tree for Saxo/Linkbricks-Horizon-AI-Korean-Gemma-2-sft-dpo-27B

Base model

google/gemma-2-27b
Quantized
(53)
this model

Datasets used to train Saxo/Linkbricks-Horizon-AI-Korean-Gemma-2-sft-dpo-27B