beomi/Yi-Ko-34B

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 34B 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-34B will come in a range of parameter sizes — 6B and 34B — with Ko(Korean+English).

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	Train tokens (per batch)
Yi-Ko-34B	A mix of Korean + English online data	34B	4k	O	40B+	5e-5	4M

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 Benchmarks

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
kmmlu_direct	N/A	none	5	exact_match	0.5027	±	0.1019
kobest_boolq	1	none	5	acc	0.9202	±	0.0072
		none	5	f1	0.9202	±	N/A
kobest_copa	1	none	5	acc	0.8480	±	0.0114
		none	5	f1	0.8479	±	N/A
kobest_hellaswag	1	none	5	acc	0.5320	±	0.0223
		none	5	f1	0.5281	±	N/A
		none	5	acc_norm	0.6340	±	0.0216
kobest_sentineg	1	none	5	acc	0.9874	±	0.0056
		none	5	f1	0.9874	±	N/A
haerae	N/A	none	5	acc	0.7965	±	0.0116
		none	5	acc_norm	0.7965	±	0.0116
- haerae_general_knowledge	1	none	5	acc	0.5114	±	0.0378
		none	5	acc_norm	0.5114	±	0.0378
- haerae_history	1	none	5	acc	0.8511	±	0.0260
		none	5	acc_norm	0.8511	±	0.0260
- haerae_loan_word	1	none	5	acc	0.8402	±	0.0283
		none	5	acc_norm	0.8402	±	0.0283
- haerae_rare_word	1	none	5	acc	0.8642	±	0.0170
		none	5	acc_norm	0.8642	±	0.0170
- haerae_standard_nomenclature	1	none	5	acc	0.8301	±	0.0305
		none	5	acc_norm	0.8301	±	0.0305

LICENSE

Follows Yi License

Citation

Acknowledgement

The training is supported by TPU Research Cloud program.