Update @ 2024.01.29 New Model beomi/Yi-Ko-DUS-9B Released! 🎉

Update @ 2023.12.03 Yi-Ko(KoEN)-6B Achieved #1🥇 Pretrained Models at Open Korean LLM Leaderboard! 🎉

Update @ 2023.12.01 Alpha Release of Yi-Ko(KoEN)-6B model 🎉

beomi/Yi-Ko-6B

Yi-Ko series models serve as advanced iterations of 01-ai/Yi models, benefiting from an expanded vocabulary and the inclusion of Korean/English corpus in its further pretraining. Just like its predecessor, Yi-Ko series models operate within the broad range of generative text models that stretch from 6 billion to 34 billion parameters. This repository focuses on the 6B pretrained version, which is tailored to fit the Hugging Face Transformers format. For access to the other models, feel free to consult the index provided below.

Model Details

Model Developers Junbum Lee (Beomi)

Variations Yi-Ko series will come in a range of parameter sizes — 6B and 34B variations.

Input Models input text only.

Output Models generate text only.

Model Architecture

Yi-Ko series models are an auto-regressive language model that uses an optimized transformer architecture based on Llama-2*.

*Yi model architecture is based on Llama2, so it can be loaded via LlamaForCausalLM class on HF.

Model Name	Training Data	Params	Context Length	GQA	Trained Tokens	LR	Batch Size(per step)
Yi-Ko-6B	A mix of Korean + English online data	6B	4k	O	>60B	5e-5	2048

Vocab Expansion

Model Name	Vocabulary Size	Description
Original Yi-Series	64000	Sentencepiece BPE
Expanded Yi-Ko Series	78464	Sentencepiece BPE. Added Korean vocab and merges

Tokenizing "안녕하세요, 오늘은 날씨가 좋네요.ㅎㅎ"

Model	# of tokens	Tokens
Original Yi-Series	47	['<0xEC>', '<0x95>', '<0x88>', '<0xEB>', '<0x85>', '<0x95>', '하', '<0xEC>', '<0x84>', '<0xB8>', '<0xEC>', '<0x9A>', '<0x94>', ',', '▁', '<0xEC>', '<0x98>', '<0xA4>', '<0xEB>', '<0x8A>', '<0x98>', '은', '▁', '<0xEB>', '<0x82>', '<0xA0>', '<0xEC>', '<0x94>', '<0xA8>', '가', '▁', '<0xEC>', '<0xA2>', '<0x8B>', '<0xEB>', '<0x84>', '<0xA4>', '<0xEC>', '<0x9A>', '<0x94>', '.', '<0xE3>', '<0x85>', '<0x8E>', '<0xE3>', '<0x85>', '<0x8E>']
Expanded Yi-Ko Series	10	['▁안녕', '하세요', ',', '▁오늘은', '▁날', '씨가', '▁좋네요', '.', 'ㅎ', 'ㅎ']
*Equal Korean vocab with Llama-2-Ko Series

Tokenizing "Llama 2: Open Foundation and Fine-Tuned Chat Models"

Model	# of tokens	Tokens
Original Yi-Series	21	['The', '▁Y', 'i', '▁series', '▁models', '▁are', '▁large', '▁language', '▁models', '▁trained', '▁from', '▁scratch', '▁by', '▁developers', '▁at', '▁', '0', '1', '.', 'AI', '.']
Expanded Yi-Ko Series	21	['▁The', '▁Y', 'i', '▁series', '▁models', '▁are', '▁large', '▁language', '▁models', '▁trained', '▁from', '▁scratch', '▁by', '▁developers', '▁at', '▁', '0', '1', '.', 'AI', '.']
*Equal Korean vocab with Llama-2-Ko Series		*Since Expanded Yi-Ko Series prepends _ at the beginning of the text(to ensure same tokenization for Korean sentences), it shows negilible difference for the first token on English tokenization.

Model Benchmark

LM Eval Harness - Korean (polyglot branch)

beomi/Yi-Ko-6B	0	5	10	50
kobest_boolq (macro_f1)	0.705806	0.79905	0.814299	0.81704
kobest_copa (macro_f1)	0.775604	0.808899	0.816866	0.842943
kobest_hellaswag (macro_f1)	0.500876	0.498673	0.493507	0.492183
kobest_sentineg (macro_f1)	0.404371	0.967254	0.982368	0.974811
kohatespeech (macro_f1)	0.353428	0.351804	0.402423	0.503764
kohatespeech_apeach (macro_f1)	0.337667	0.498679	0.471962	0.608401
kohatespeech_gen_bias (macro_f1)	0.124535	0.484745	0.474475	0.461714
korunsmile (f1)	0.382804	0.349344	0.391383	0.432875
nsmc (acc)	0.55064	0.8801	0.89866	0.9071
pawsx_ko (acc)	0.5145	0.54	0.538	0.5165

LICENSE

Apache 2.0 (for research)

For commercial purpose, mailto: [email protected] to acquire Yi-Ko sereis commercial license.

Citation

Please use this bibtex below:

@misc {lee_junbum_2024,
    author       = { {Lee Junbum} },
    title        = { Yi-Ko-6B (Revision 205083a) },
    year         = 2024,
    url          = { https://huggingface.co/beomi/Yi-Ko-6B },
    doi          = { 10.57967/hf/1708 },
    publisher    = { Hugging Face }
}

Acknowledgement

The training is supported by TPU Research Cloud program.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	50.27
AI2 Reasoning Challenge (25-Shot)	48.89
HellaSwag (10-Shot)	74.48
MMLU (5-Shot)	55.72
TruthfulQA (0-shot)	37.09
Winogrande (5-shot)	72.93
GSM8k (5-shot)	12.51