(주)미디어그룹사람과숲과 (주)마커의 LLM 연구 컨소시엄에서 개발된 모델입니다
The license is cc-by-nc-sa-4.0.
KoT-platypus2
CoT + KO-platypus2 = KoT-platypus2
Model Details
Model Developers Kyujin Han (kyujinpy)
Input Models input text only.
Output Models generate text only.
Model Architecture
KoT-platypus2-13B is an auto-regressive language model based on the LLaMA2 transformer architecture.
Repo Link
Github KoT-platypus: KoT-platypus2
Base Model
KO-Platypus2-13B
More detail repo(Github): CoT-llama2
More detail repo(Github): KO-Platypus2
Training Dataset
I use KoCoT_2000.
Using DeepL, translate about kaist-CoT.
I use A100 GPU 40GB and COLAB, when trianing.
Training Hyperparameters
| Hyperparameters | Value |
|---|---|
| batch_size | 64 |
| micro_batch_size | 1 |
| Epochs | 15 |
| learning_rate | 1e-5 |
| cutoff_len | 4096 |
| lr_scheduler | linear |
| base_model | kyujinpy/KO-Platypus2-13B |
Model Benchmark
KO-LLM leaderboard
- Follow up as Open KO-LLM LeaderBoard.
| Model | Average | Ko-ARC | Ko-HellaSwag | Ko-MMLU | Ko-TruthfulQA | Ko-CommonGen V2 |
|---|---|---|---|---|---|---|
| KoT-Platypus2-13B(ours) | 49.55 | 43.69 | 53.05 | 42.29 | 43.34 | 65.38 |
| KO-Platypus2-13B | 47.90 | 44.20 | 54.31 | 42.47 | 44.41 | 54.11 |
| hyunseoki/ko-en-llama2-13b | 46.68 | 42.15 | 54.23 | 38.90 | 40.74 | 57.39 |
| MarkrAI/kyujin-CoTy-platypus-ko-12.8b | 46.44 | 34.98 | 49.11 | 25.68 | 37.59 | 84.86 |
| momo/polyglot-ko-12.8b-Chat-QLoRA-Merge | 45.71 | 35.49 | 49.93 | 25.97 | 39.43 | 77.70 |
Compare with Top 4 SOTA models. (update: 10/07)
Implementation Code
### KO-Platypus
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
repo = "kyujinpy/KoT-platypus2-13B"
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: beomi/llama-2-ko-7b