Update Log

2023.12.27
- New Model is here! Trained with only open-accessible Korean text corpus: https://huggingface.co/beomi/open-llama-2-ko-7b
2023.10.19
- Fix Tokenizer bug(space not applied when decoding) after transforemrs>=4.34.0

Llama-2-Ko 🦙🇰🇷

Llama-2-Ko serves as an advanced iteration of Llama 2, benefiting from an expanded vocabulary and the inclusion of a Korean corpus in its further pretraining. Just like its predecessor, Llama-2-Ko operates within the broad range of generative text models that stretch from 7 billion to 70 billion parameters. This repository focuses on the 7B 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 Llama-2-Ko will come in a range of parameter sizes — 7B, 13B, and 70B — as well as pretrained and fine-tuned variations.

Input Models input text only.

Output Models generate text only.

Model Architecture

Llama-2-Ko is an auto-regressive language model that uses an optimized transformer architecture based on Llama-2.

	Training Data	Params	Content Length	GQA	Tokens	LR
Llama 2	A new mix of Korean online data	7B	4k	✗	>40B*	1e^-5
*Plan to train upto 200B tokens

Vocab Expansion

Model Name	Vocabulary Size	Description
Original Llama-2	32000	Sentencepiece BPE
Expanded Llama-2-Ko	46336	Sentencepiece BPE. Added Korean vocab and merges

Tokenizing "안녕하세요, 오늘은 날씨가 좋네요."

Model	Tokens
Llama-2	`['▁', '안', '<0xEB>', '<0x85>', '<0x95>', '하', '세', '요', ',', '▁', '오', '<0xEB>', '<0x8A>', '<0x98>', '은', '▁', '<0xEB>', '<0x82>', '<0xA0>', '씨', '가', '▁', '<0xEC>', '<0xA2>', '<0x8B>', '<0xEB>', '<0x84>', '<0xA4>', '요']`
Llama-2-Ko	`['▁안녕', '하세요', ',', '▁오늘은', '▁날', '씨가', '▁좋네요']`

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

Model	Tokens
Llama-2	`['▁L', 'l', 'ama', '▁', '2', ':', '▁Open', '▁Foundation', '▁and', '▁Fine', '-', 'T', 'un', 'ed', '▁Ch', 'at', '▁Mod', 'els']`
Llama-2-Ko	`['▁L', 'l', 'ama', '▁', '2', ':', '▁Open', '▁Foundation', '▁and', '▁Fine', '-', 'T', 'un', 'ed', '▁Ch', 'at', '▁Mod', 'els']`

Model Benchmark

LM Eval Harness - Korean (polyglot branch)

Used EleutherAI's lm-evaluation-harness https://github.com/EleutherAI/lm-evaluation-harness/tree/polyglot

NSMC (Acc) - 50000 full test

TBD

COPA (F1)

Model	0-shot	5-shot	10-shot	50-shot
https://huggingface.co/skt/ko-gpt-trinity-1.2B-v0.5	0.6696	0.6477	0.6419	0.6514
https://huggingface.co/kakaobrain/kogpt	0.7345	0.7287	0.7277	0.7479
https://huggingface.co/facebook/xglm-7.5B	0.6723	0.6731	0.6769	0.7119
https://huggingface.co/EleutherAI/polyglot-ko-1.3b	0.7196	0.7193	0.7204	0.7206
https://huggingface.co/EleutherAI/polyglot-ko-3.8b	0.7595	0.7608	0.7638	0.7788
https://huggingface.co/EleutherAI/polyglot-ko-5.8b	0.7745	0.7676	0.7775	0.7887
https://huggingface.co/EleutherAI/polyglot-ko-12.8b	0.7937	0.8108	0.8037	0.8369
Llama-2 Original 7B*	0.562033	0.575982	0.576216	0.595532
Llama-2-Ko-7b 20B (10k)	0.738780	0.762639	0.780761	0.797863
Llama-2-Ko-7b 40B (20k)	0.743630	0.792716	0.803746	0.825944
*Llama-2 Original 7B used https://huggingface.co/meta-llama/Llama-2-7b-hf (w/o tokenizer updated)

HellaSwag (F1)

Model	0-shot	5-shot	10-shot	50-shot
https://huggingface.co/skt/ko-gpt-trinity-1.2B-v0.5	0.5243	0.5272	0.5166	0.5352
https://huggingface.co/kakaobrain/kogpt	0.5590	0.5833	0.5828	0.5907
https://huggingface.co/facebook/xglm-7.5B	0.5665	0.5689	0.5565	0.5622
https://huggingface.co/EleutherAI/polyglot-ko-1.3b	0.5247	0.5260	0.5278	0.5427
https://huggingface.co/EleutherAI/polyglot-ko-3.8b	0.5707	0.5830	0.5670	0.5787
https://huggingface.co/EleutherAI/polyglot-ko-5.8b	0.5976	0.5998	0.5979	0.6208
https://huggingface.co/EleutherAI/polyglot-ko-12.8b	0.5954	0.6306	0.6098	0.6118
Llama-2 Original 7B*	0.415390	0.431382	0.421342	0.442003
Llama-2-Ko-7b 20B (10k)	0.451757	0.466751	0.472607	0.482776
Llama-2-Ko-7b 40B (20k)	0.456246	0.465665	0.469810	0.477374
*Llama-2 Original 7B used https://huggingface.co/meta-llama/Llama-2-7b-hf (w/o tokenizer updated)

BoolQ (F1)

Model	0-shot	5-shot	10-shot	50-shot
https://huggingface.co/skt/ko-gpt-trinity-1.2B-v0.5	0.3356	0.4014	0.3640	0.3560
https://huggingface.co/kakaobrain/kogpt	0.4514	0.5981	0.5499	0.5202
https://huggingface.co/facebook/xglm-7.5B	0.4464	0.3324	0.3324	0.3324
https://huggingface.co/EleutherAI/polyglot-ko-1.3b	0.3552	0.4751	0.4109	0.4038
https://huggingface.co/EleutherAI/polyglot-ko-3.8b	0.4320	0.5263	0.4930	0.4038
https://huggingface.co/EleutherAI/polyglot-ko-5.8b	0.4356	0.5698	0.5187	0.5236
https://huggingface.co/EleutherAI/polyglot-ko-12.8b	0.4818	0.6041	0.6289	0.6448
Llama-2 Original 7B*	0.352050	0.563238	0.474788	0.419222
Llama-2-Ko-7b 20B (10k)	0.360656	0.679743	0.680109	0.662152
Llama-2-Ko-7b 40B (20k)	0.578640	0.697747	0.708358	0.714423
*Llama-2 Original 7B used https://huggingface.co/meta-llama/Llama-2-7b-hf (w/o tokenizer updated)

SentiNeg (F1)

Model	0-shot	5-shot	10-shot	50-shot
https://huggingface.co/skt/ko-gpt-trinity-1.2B-v0.5	0.6065	0.6878	0.7280	0.8413
https://huggingface.co/kakaobrain/kogpt	0.3747	0.8942	0.9294	0.9698
https://huggingface.co/facebook/xglm-7.5B	0.3578	0.4471	0.3964	0.5271
https://huggingface.co/EleutherAI/polyglot-ko-1.3b	0.6790	0.6257	0.5514	0.7851
https://huggingface.co/EleutherAI/polyglot-ko-3.8b	0.4858	0.7950	0.7320	0.7851
https://huggingface.co/EleutherAI/polyglot-ko-5.8b	0.3394	0.8841	0.8808	0.9521
https://huggingface.co/EleutherAI/polyglot-ko-12.8b	0.9117	0.9015	0.9345	0.9723
Llama-2 Original 7B*	0.347502	0.529124	0.480641	0.788457
Llama-2-Ko-7b 20B (10k)	0.485546	0.829503	0.871141	0.851253
Llama-2-Ko-7b 40B (20k)	0.459447	0.761079	0.727611	0.936988
*Llama-2 Original 7B used https://huggingface.co/meta-llama/Llama-2-7b-hf (w/o tokenizer updated)

Note for oobabooga/text-generation-webui

Remove ValueError at load_tokenizer function(line 109 or near), in modules/models.py.

diff --git a/modules/models.py b/modules/models.py
index 232d5fa..de5b7a0 100644
--- a/modules/models.py
+++ b/modules/models.py
@@ -106,7 +106,7 @@ def load_tokenizer(model_name, model):
                 trust_remote_code=shared.args.trust_remote_code,
                 use_fast=False
             )
-        except ValueError:
+        except:
             tokenizer = AutoTokenizer.from_pretrained(
                 path_to_model,
                 trust_remote_code=shared.args.trust_remote_code,

Since Llama-2-Ko uses FastTokenizer provided by HF tokenizers NOT sentencepiece package, it is required to use use_fast=True option when initialize tokenizer.

Apple Sillicon does not support BF16 computing, use CPU instead. (BF16 is supported when using NVIDIA GPU)

Citation

@misc {l._junbum_2023,
    author       = { {L. Junbum} },
    title        = { llama-2-ko-7b (Revision 4a9993e) },
    year         = 2023,
    url          = { https://huggingface.co/beomi/llama-2-ko-7b },
    doi          = { 10.57967/hf/1098 },
    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.	39.43
ARC (25-shot)	48.46
HellaSwag (10-shot)	75.28
MMLU (5-shot)	39.56
TruthfulQA (0-shot)	34.49
Winogrande (5-shot)	72.14
GSM8K (5-shot)	1.97
DROP (3-shot)	4.1