kyujinpy/Kosy-platypus2-13B-v4

Kosy🍵llama

Model Details

Model Developers Kyujin Han (kyujinpy)

Model Description
NEFTune method를 활용하여 훈련한 Ko-platypus2 new version!
(Noisy + KO + llama = Kosy🍵llama)

Repo Link
Github KoNEFTune: Kosy🍵llama
If you visit our github, you can easily apply Random_noisy_embedding_fine-tuning!!

Base Model
hyunseoki/ko-en-llama2-13b

Training Dataset
Version of combined dataset: kyujinpy/KOpen-platypus
I use A100 GPU 40GB and COLAB, when trianing.

Model comparisons

KO-LLM leaderboard

NEFT comparisons

Model	Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
Ko-Platypus2-13B	45.60	44.20	54.31	42.47	44.41	42.62
*NEFT(🍵kosy)+MLP-v1	43.64	43.94	53.88	42.68	43.46	34.24
*NEFT(🍵kosy)+MLP-v2	45.45	44.20	54.56	42.60	42.68	42.98
*NEFT(🍵kosy)+MLP-v3	46.31	43.34	54.54	43.38	44.11	46.16
NEFT(🍵kosy)+Attention	44.92	42.92	54.48	42.99	43.00	41.20
NEFT(🍵kosy)	45.08	43.09	53.61	41.06	43.47	43.21

*Different Hyperparameters such that learning_rate, batch_size, epoch, etc...

Implementation Code

### KO-Platypus
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "kyujinpy/Koisy-Platypus2-13B"
OpenOrca = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)

---