Model	Tokens
Platypus2-7b	`['▁', '안', '<0xEB>', '<0x85>', '<0x95>', '하', '세', '요', ',', '▁', '오', '<0xEB>', '<0x8A>', '<0x98>', '은', '▁', '<0xEB>', '<0x82>', '<0xA0>', '씨', '가', '▁', '<0xEC>', '<0xA2>', '<0x8B>', '<0xEB>', '<0x84>', '<0xA4>', '요', '.']`
KO-Platypus2-7b-ex	`['▁안녕', '하세요', ',', '▁오늘은', '▁날', '씨가', '▁좋네요', '.']`

Tokenizing "Platypus: Quick, Cheap, and Powerful Refinement of LLMs"

Model	Tokens
Platypus2-7b	`['▁Plat', 'yp', 'us', ':', '▁Quick', ',', '▁Che', 'ap', ',', '▁and', '▁Power', 'ful', '▁Re', 'fin', 'ement', '▁of', '▁L', 'LM', 's']`
KO-Platypus2-7b-ex	`[▁Plat', 'yp', 'us', ':', '▁Quick', ',', '▁Che', 'ap', ',', '▁and', '▁Power', 'ful', '▁Re', 'fin', 'ement', '▁of', '▁L', 'LM', 's']`

Model Benchmark

LM Eval Harness - Korean (polyglot branch)

Used EleutherAI's lm-evaluation-harness

Question Answering (QA)

COPA (F1)

Model	0-shot	5-shot	10-shot	50-shot
Polyglot-ko-1.3b	0.7196	0.7193	0.7204	0.7206
Polyglot-ko-3.8b	0.7595	0.7608	0.7638	0.7788
Polyglot-ko-5.8b	0.7745	0.7676	0.7775	0.7887
Polyglot-ko-12.8b	0.7937	0.8108	0.8037	0.8369
Llama-2-Ko-7b 20B	0.7388	0.7626	0.7808	0.7979
Llama-2-Ko-7b 40B	0.7436	0.7927	0.8037	0.8259
*Platypus2-7B	0.5594	0.5913	0.5863	0.5916
KO-platypus2-7B-EX(ours)	0.7509	0.7899	0.8029	0.8290
*Platypus2-7B Original used https://huggingface.co/garage-bAInd/Platypus2-7B

Natural Language Inference (NLI; 자연어 추론 평가)

HellaSwag (F1)

Model	0-shot	5-shot	10-shot	50-shot
Polyglot-ko-1.3b	0.5247	0.5260	0.5278	0.5427
Polyglot-ko-3.8b	0.5707	0.5830	0.5670	0.5787
Polyglot-ko-5.8b	0.5976	0.5998	0.5979	0.6208
Polyglot-ko-12.8b	0.5954	0.6306	0.6098	0.6118
Llama-2-Ko-7b 20B	0.4518	0.4668	0.4726	0.4828
Llama-2-Ko-7b 40B	0.4562	0.4657	0.4698	0.4774
*Platypus2-7B	0.4097	0.4258	0.4358	0.4271
KO-platypus2-7B-EX(ours)	0.4571	0.4461	0.4371	0.4525
*Platypus2-7B Original used https://huggingface.co/garage-bAInd/Platypus2-7B

Question Answering (QA)

BoolQ (F1)

Model	0-shot	5-shot	10-shot	50-shot
Polyglot-ko-1.3b	0.3552	0.4751	0.4109	0.4038
Polyglot-ko-3.8b	0.4320	0.5263	0.4930	0.4038
Polyglot-ko-5.8b	0.4356	0.5698	0.5187	0.5236
Polyglot-ko-12.8b	0.4818	0.6041	0.6289	0.6448
Llama-2-Ko-7b 20B	0.3607	0.6797	0.6801	0.6622
Llama-2-Ko-7b 40B	0.5786	0.6977	0.7084	0.7144
*Platypus2-7B	0.3419	0.6024	0.5630	0.5461
KO-platypus2-7B-EX(ours)	0.6028	0.6979	0.7016	0.6988
*Platypus2-7B Original used https://huggingface.co/garage-bAInd/Platypus2-7B

Classification

SentiNeg (F1)

Model	0-shot	5-shot	10-shot	50-shot
Polyglot-ko-1.3b	0.6790	0.6257	0.5514	0.7851
Polyglot-ko-3.8b	0.4858	0.7950	0.7320	0.7851
Polyglot-ko-5.8b	0.3394	0.8841	0.8808	0.9521
Polyglot-ko-12.8b	0.9117	0.9015	0.9345	0.9723
Llama-2-Ko-7b 20B	0.4855	0.8295	0.8711	0.8513
Llama-2-Ko-7b 40B	0.4594	0.7611	0.7276	0.9370
*Platypus2-7B	0.4098	0.7388	0.7558	0.8129
KO-platypus2-7B-EX(ours)	0.5821	0.7653	0.7991	0.8643
*Platypus2-7B Original used https://huggingface.co/garage-bAInd/Platypus2-7B

Implementation Code

python

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

repo = "kyujinpy/KO-Platypus2-7B-ex"
ko_platypus = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
ko_platypus_tokenizer = AutoTokenizer.from_pretrained(repo)

Readme format: beomi/llama-2-ko-7b

Below is the original model card of the Platypus2-13B model.

Platypus2-13B

Platypus-13B is an instruction fine-tuned model based on the LLaMA2-13B transformer architecture.

Benchmark Metrics

Metric	Value
MMLU (5-shot)	56.70
ARC (25-shot)	61.26
HellaSwag (10-shot)	82.56
TruthfulQA (0-shot)	44.86
Avg.	61.35

We use state-of-the-art Language Model Evaluation Harness to run the benchmark tests above, using the same version as the HuggingFace LLM Leaderboard. Please see below for detailed instructions on reproducing benchmark results.

Model Details

Trained by: Cole Hunter & Ariel Lee
Model type: Platypus2-13B is an auto-regressive language model based on the LLaMA2 transformer architecture.
Language(s): English
License for base weights: Non-Commercial Creative Commons license (CC BY-NC-4.0)

Prompt Template

### Instruction:

<prompt> (without the <>)

### Response:

Training Dataset

garage-bAInd/Platypus2-13B trained using STEM and logic based dataset garage-bAInd/Open-Platypus.

Please see our paper and project webpage for additional information.

Training Procedure

garage-bAInd/Platypus2-13B was instruction fine-tuned using LoRA on 1 A100 80GB. For training details and inference instructions please see the Platypus2 GitHub repo.

Reproducing Evaluation Results

Install LM Evaluation Harness:

# clone repository
git clone https://github.com/EleutherAI/lm-evaluation-harness.git
# check out the correct commit
git checkout b281b0921b636bc36ad05c0b0b0763bd6dd43463
# change to repo directory
cd lm-evaluation-harness
# install
pip install -e .

Each task was evaluated on 1 A100 80GB GPU.

ARC:

python main.py --model hf-causal-experimental --model_args pretrained=garage-bAInd/Platypus2-13B --tasks arc_challenge --batch_size 1 --no_cache --write_out --output_path results/Platypus2-13B/arc_challenge_25shot.json --device cuda --num_fewshot 25

HellaSwag:

python main.py --model hf-causal-experimental --model_args pretrained=garage-bAInd/Platypus2-13B --tasks hellaswag --batch_size 1 --no_cache --write_out --output_path results/Platypus2-13B/hellaswag_10shot.json --device cuda --num_fewshot 10

MMLU:

python main.py --model hf-causal-experimental --model_args pretrained=garage-bAInd/Platypus2-13B --tasks hendrycksTest-* --batch_size 1 --no_cache --write_out --output_path results/Platypus2-13B/mmlu_5shot.json --device cuda --num_fewshot 5

TruthfulQA:

python main.py --model hf-causal-experimental --model_args pretrained=garage-bAInd/Platypus2-13B --tasks truthfulqa_mc --batch_size 1 --no_cache --write_out --output_path results/Platypus2-13B/truthfulqa_0shot.json --device cuda

Limitations and bias

Llama 2 and fine-tuned variants are a new technology that carries risks with use. Testing conducted to date has been in English, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, Llama 2 and any fine-tuned varient's potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of Llama 2 variants, developers should perform safety testing and tuning tailored to their specific applications of the model.

Please see the Responsible Use Guide available at https://ai.meta.com/llama/responsible-use-guide/

Citations

bibtex

@article{platypus2023,
    title={Platypus: Quick, Cheap, and Powerful Refinement of LLMs}, 
    author={Ariel N. Lee and Cole J. Hunter and Nataniel Ruiz},
    booktitle={arXiv preprint arxiv:2308.07317},
    year={2023}
}

bibtex

@misc{touvron2023llama,
    title={Llama 2: Open Foundation and Fine-Tuned Chat Models}, 
    author={Hugo Touvron and Louis Martin and Kevin Stone and Peter Albert and Amjad Almahairi and Yasmine Babaei and Nikolay Bashlykov       year={2023},
    eprint={2307.09288},
    archivePrefix={arXiv},
}

bibtex

@inproceedings{
    hu2022lora,
    title={Lo{RA}: Low-Rank Adaptation of Large Language Models},
    author={Edward J Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=nZeVKeeFYf9}
}