# pip install accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("MDDDDR/Ko-Luxia-8B-it-v0.1")
model = AutoModelForCausalLM.from_pretrained(
    "MDDDDR/Ko-Luxia-8B-it-v0.1",
    device_map="auto",
    torch_dtype=torch.bfloat16
)

input_text = "사과가 뭐야?"
input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Training dataset

dataset : kyujinpy/KOpen-platypus

lora_config and bnb_config in Training

python

bnd_config = BitsAndBytesConfig(
  load_in_4bit = True,
  bnb_4bit_use_double_quant = True,
  bnb_4bit_quant_type = 'nf4',
  bnb_4bit_compute_dtype = torch.bfloat16
)

lora_config = LoraConfig(
  r = 16,
  lora_alpha = 16,
  lora_dropout = 0.05,
  target_modules = ['gate_proj', 'up_proj', 'down_proj']
)

Hardware

RTX 3090 Ti 24GB x 1

Evaluation Benchmark Results

Tasks	Version	Filter	Metric		Value		Stderr
kobest_boolq	1	none	acc	↑	0.6425	±	0.0128
		none	f1	↑	0.6054	±	N/A
kobest_copa	1	none	acc	↑	0.7340	±	0.0140
		none	f1	↑	0.7333	±	N/A
kobest_hellaswag	1	none	acc	↑	0.4760	±	0.0224
		none	acc_norm	↑	0.6120	±	0.0218
		none	f1	↑	0.4745	±	N/A
kobest_sentineg	1	none	acc	↑	0.5894	±	0.0247
		none	f1	↑	0.5682	±	N/A