Custom-KoLLM-13B-v4 / README.md
kyujinpy's picture
Upload README.md
84e8fe3
metadata
language:
  - ko
datasets:
  - kyujinpy/KOR-OpenOrca-Platypus-v3
library_name: transformers
pipeline_tag: text-generation
license: cc-by-nc-sa-4.0

⭐My custom LLM 13B⭐

Model Details

Model Developers

  • Kyujin Han (kyujinpy)

Model Architecture

  • My custom LLM 13B is an auto-regressive language model based on the LLaMA2 transformer architecture.

Base Model

Training Dataset


Model comparisons

Ko-LLM leaderboard(11/27; link)

Model Average Ko-ARC Ko-HellaSwag Ko-MMLU Ko-TruthfulQA Ko-CommonGen V2
⭐My custom LLM 13B-v1⭐ 50.19 45.99 56.93 41.78 41.66 64.58
⭐My custom LLM 13B-v2⭐ 48.28 45.73 56.97 38.77 38.75 61.16
⭐My custom LLM 13B-v4⭐ 49.89 45.05 57.06 41.83 42.93 62.57

Model comparisons2

AI-Harness evaluation; link

Model Copa Copa HellaSwag HellaSwag BoolQ BoolQ Sentineg Sentineg
0-shot 5-shot 0-shot 5-shot 0-shot 5-shot 0-shot 5-shot
⭐My custom LLM 13B-v1⭐ 0.7987 0.8269 0.4994 0.5660 0.3343 0.5060 0.6984 0.9723
⭐My custom LLM 13B-v2⭐ 0.7938 0.8209 0.4978 0.4893 0.3343 0.5614 0.6283 0.9773
⭐My custom LLM 13B-v4⭐ 0.7988 0.8279 0.4995 0.4953 0.3343 0.3558 0.7825 0.9698
beomi/llama-2-koen-13b 0.7768 0.8128 0.4999 0.5127 0.3988 0.7038 0.5870 0.9748

Implementation Code

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

repo = "PracticeLLM/Custom-KoLLM-13B-v4"
OpenOrca = AutoModelForCausalLM.from_pretrained(
        repo,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map='auto'
)
OpenOrca_tokenizer = AutoTokenizer.from_pretrained(repo)

Hyperparameters

  • learning_rate: 4e-4
  • batch_size: 16
  • epoch: 1
  • lora_target_modules: [gate_proj, down_proj, up_proj, q_proj, k_proj, v_proj, o_proj]
  • cutoff_len: 4096