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

tokenizer = AutoTokenizer.from_pretrained("MDDDDR/Meta-Llama-3.1-8B-it-v0.1")
model = AutoModelForCausalLM.from_pretrained(
    "MDDDDR/Meta-Llama-3.1-8B-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_1 : DILAB-HYU/KoQuality

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 = 8,
  lora_alpha = 8,
  lora_dropout = 0.05,
  target_modules = ['gate_proj', 'up_proj', 'down_proj']
)

Model evaluation

Tasks	Version	Filter	Metric		Value		Stderr
kobest_boolq	1	none	acc	↑	0.5150	±	0.0133
		none	f1	↑	0.3634	±	N/A
kobest_copa	1	none	acc	↑	0.6280	±	0.0153
		none	f1	↑	0.6279	±	N/A
kobest_hellaswag	1	none	acc	↑	0.4280	±	0.0221
		none	acc_norm	↑	0.5540	±	0.0223
		none	f1	↑	0.4250	±	N/A
kobest_sentineg	1	none	acc	↑	0.7406	±	0.0220
		none	f1	↑	0.7317	±	N/A

Hardware

RTX 3090 Ti 24GB x 1
Training Time : 1 hours