MDDDDR/gemma-7B-it-v0.2

Model Card for Model ID

base_model : google/gemma-7b

Basic usage

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

tokenizer = AutoTokenizer.from_pretrained("MDDDDR/gemma-7b-it-v0.2")
model = AutoModelForCausalLM.from_pretrained(
    "MDDDDR/gemma-7b-it-v0.2",
    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 : MarkrAI/KoCommercial-Dataset
• dataset_2 : nlpai-lab/kullm-v2

lora_config and bnb_config in Training

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	n-shot	Metric		Value		Stderr
kobest_boolq	1	none	0	acc	↑	0.5912	±	0.0131
		none	0	f1	↑	0.5183	±	N/A
kobest_copa	1	none	0	acc	↑	0.6320	±	0.0153
		none	0	f1	↑	0.6313	±	N/A
kobest_hellaswag	1	none	0	acc	↑	0.4220	±	0.0221
		none	0	acc_norm	↑	0.5280	±	0.0223
		none	0	f1	↑	0.4190	±	N/A
kobest_sentineg	1	none	0	acc	↑	0.4962	±	0.0251
		none	0	f1	↑	0.3747	±	N/A

Hardware

• RTX 3090 Ti 24GB x 1
• Training Time : 80 hours