library_name: transformers
license: apache-2.0
base_model: mistralai/Mistral-Small-Instruct-2409
datasets:
- Saxo/ko_cn_translation_tech_social_science_linkbricks_single_dataset
- Saxo/ko_jp_translation_tech_social_science_linkbricks_single_dataset
- >-
Saxo/en_ko_translation_tech_science_linkbricks_single_dataset_with_prompt_text_huggingface
- >-
Saxo/en_ko_translation_social_science_linkbricks_single_dataset_with_prompt_text_huggingface
- >-
Saxo/ko_aspect_sentiment_sns_mall_sentiment_linkbricks_single_dataset_with_prompt_text_huggingface
- Saxo/ko_summarization_linkbricks_single_dataset_with_prompt_text_huggingface
- >-
Saxo/OpenOrca_cleaned_kor_linkbricks_single_dataset_with_prompt_text_huggingface
- >-
Saxo/ko_government_qa_total_linkbricks_single_dataset_with_prompt_text_huggingface_sampled
- Saxo/ko-news-corpus-1
- Saxo/ko-news-corpus-2
- Saxo/ko-news-corpus-3
- Saxo/ko-news-corpus-4
- Saxo/ko-news-corpus-5
- Saxo/ko-news-corpus-6
- Saxo/ko-news-corpus-7
- Saxo/ko-news-corpus-8
- Saxo/ko-news-corpus-9
- maywell/ko_Ultrafeedback_binarized
- youjunhyeok/ko-orca-pair-and-ultrafeedback-dpo
- lilacai/glaive-function-calling-v2-sharegpt
- kuotient/gsm8k-ko
language:
- ko
- en
- jp
- cn
pipeline_tag: text-generation
Model Card for Model ID
AI ์ ๋น
๋ฐ์ดํฐ ๋ถ์ ์ ๋ฌธ ๊ธฐ์
์ธ Linkbricks์ ๋ฐ์ดํฐ์ฌ์ด์ธํฐ์คํธ์ธ ์ง์ค์ฑ(Saxo) ๋ฐ์ฌ๊ฐ
Mistral-Small-Instruct-2409 ๋ฒ ์ด์ค๋ชจ๋ธ์ ์ฌ์ฉํด์ H100-80G 8๊ฐ๋ฅผ ํตํด ์ฝ 40%์ ๋์ ํ๋ผ๋ฏธํฐ๋ฅผ ํ๊ตญ์ด CPT(Continued-Pretraining)->SFT->DPO ํ ํ๊ธ ์ธ์ด ๋ชจ๋ธ
9์ฒ๋ง๊ฑด์ ํ๊ธ ๋ด์ค ์ฝํผ์ค๋ฅผ ๊ธฐ์ค์ผ๋ก ๋ค์ํ ํ
์คํฌ๋ณ ํ๊ตญ์ด-์ค๊ตญ์ด-์์ด-์ผ๋ณธ์ด ๊ต์ฐจ ํ์ต ๋ฐ์ดํฐ์ ์ํ ๋ฐ ๋
ผ๋ฆฌํ๋จ ๋ฐ์ดํฐ๋ฅผ ํตํ์ฌ ํ์ค์ผ์ ์ธ์ด ๊ต์ฐจ ์ฆ๊ฐ ์ฒ๋ฆฌ์ ๋ณต์กํ ๋
ผ๋ฆฌ ๋ฌธ์ ์ญ์ ๋์ ๊ฐ๋ฅํ๋๋ก ํ๋ จํ ๋ชจ๋ธ์ด๋ค.
-ํ ํฌ๋์ด์ ๋ ๋จ์ด ํ์ฅ ์์ด ๋ฒ ์ด์ค ๋ชจ๋ธ ๊ทธ๋๋ก ์ฌ์ฉ
-๊ณ ๊ฐ ๋ฆฌ๋ทฐ๋ ์์
ํฌ์คํ
๊ณ ์ฐจ์ ๋ถ์ ๋ฐ ์ฝ๋ฉ๊ณผ ์๋ฌธ, ์ํ, ๋
ผ๋ฆฌํ๋จ ๋ฑ์ด ๊ฐํ๋ ๋ชจ๋ธ
-32k ์ํ์ค ๊ธธ์ด
-ํ์
์ฝ ์ง์
-Deepspeed Stage=3, rslora ๋ฐ BAdam Layer Mode ์ฌ์ฉ
Finetuned by Mr. Yunsung Ji (Saxo), a data scientist at Linkbricks, a company specializing in AI and big data analytics
about 40% of total parameters Korean CPT(Continued-Pretraining)->SFT->DPO training model based on Mistral-Small-Instruct-2409 through 8 H100-80Gs as a Korean language model
It is a model that has been trained to handle Korean-Chinese-English-Japanese cross-training data and 90M korean news corpus and logic judgment data for various tasks to enable cross-fertilization processing and complex Korean logic & math problems.
-Tokenizer uses the base model without word expansion
-Models enhanced with high-dimensional analysis of customer reviews and social posts, as well as coding, writing, math and decision making
-32k sequence length
-Function calling
-Deepspeed Stage=3, use rslora and BAdam Layer Mode