README.md

metadata

language:
  - ko
datasets:
  - kyujinpy/KOpen-platypus
library_name: transformers
pipeline_tag: text-generation
license: cc-by-nc-sa-4.0

(주)미디어그룹사람과숲과 (주)마커의 LLM 연구 컨소시엄에서 개발된 모델입니다
The license is cc-by-nc-sa-4.0.

Poly-platypus-ko

Polyglot-ko + KO-platypus2 = Poly-platypus-ko

Model Details

Model Developers Kyujin Han (kyujinpy)
Input Models input text only.
Output Models generate text only.
Model Architecture
Poly-platypus-ko is an auto-regressive language model based on the polyglot-ko transformer architecture.

Repo Link
Github KO-platypus2: KO-platypus2
Github Poly-platypus-ko: Poly-platypus-ko

Base Model
Polyglot-ko-12.8b

Fine-tuning method
Same as KO-Platypus2.

Training Dataset
I use KOpen-platypus dataset.
I use A100 GPU 40GB and COLAB, when trianing.

Model Bechmark1

KO-LLM leaderboard

• Follow up as Open KO-LLM LeaderBoard.

Model	Average	Ko-ARC	Ko-HellaSwag	Ko-MMLU	Ko-TruthfulQA	Ko-CommonGen V2
Poly-platypus-ko-12.8b(ours)	44.95	35.15	50.39	25.58	38.74	74.88
KoT-platypus2-7B	45.62	38.05	49.63	34.68	37.69	68.08
KO-platypus2-7B-EX	45.41	39.08	50.86	34.60	37.94	64.55
42MARU/polyglot-ko-12.8b-instruct	43.89	36.35	51.59	26.38	45.16	59.98
FINDA-FIT/llama-p	43.63	39.59	50.74	33.85	38.09	55.87

Compare with Top 4 SOTA models. (update: 10/01)

---

Model Benchmark2

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-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
KO-platypus2-7B-EX	0.7509	0.7899	0.8029	0.8290
KoT-platypus2-7B	0.7517	0.7868	0.8009	0.8239
Poly-platypus-ko-12.8b(ours)	0.7876	0.8099	0.8008	0.8239

Natural Language Inference (NLI; 자연어 추론 평가)

HellaSwag (F1)

Model	0-shot	5-shot	10-shot	50-shot
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
KO-platypus2-7B-EX	0.4571	0.4461	0.4371	0.4525
KoT-platypus2-7B	0.4432	0.4382	0.4550	0.4534
Poly-platypus-ko-12.8b(ours)	0.4838	0.4858	0.5005	0.5062

Question Answering (QA)

BoolQ (F1)

Model	0-shot	5-shot	10-shot	50-shot
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
KO-platypus2-7B-EX	0.6028	0.6979	0.7016	0.6988
KoT-platypus2-7B	0.6142	0.6757	0.6839	0.6878
Poly-platypus-ko-12.8b(ours)	0.4888	0.6520	0.6568	0.6835

Classification

SentiNeg (F1)

Model	0-shot	5-shot	10-shot	50-shot
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
KO-platypus2-7B-EX	0.5821	0.7653	0.7991	0.8643
KoT-platypus2-7B	0.6127	0.7199	0.7531	0.8381
Poly-platypus-ko-12.8b(ours)	0.8490	0.9597	0.9723	0.9847

Implementation Code

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

repo = "MarkrAI/kyujin-Poly-platypus-ko-12.8b"
CoT-llama = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
CoT-llama_tokenizer = AutoTokenizer.from_pretrained(repo)

Readme format: kyujinpy/KoT-platypus2-7B

---