Parameter	Value
Base Model	meta-llama/Llama-3.1-8B-Instruct
Epochs	3
Learning Rate	3e-4
LR Scheduler	Cosine
Warmup Steps	100
Micro Batch Size	4
Gradient Accumulation Steps	32
Effective Batch Size	512 (4 GPUs x 4 x 32)
Max Sequence Length	4096
Precision	BF16

LoRA Configuration

Parameter	Value
LoRA Rank (r)	8
LoRA Alpha	16
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, up_proj, down_proj, gate_proj

Quantization (QLoRA)

4-bit quantization with NF4
Double quantization enabled
Compute dtype: BF16

Training Data

Korean SFT dataset (ko_combined_sft_dataset.json)
Prompt template: Alpaca format
Validation set: 5% of training data

Training Infrastructure

GPUs: 4x NVIDIA GPUs
Framework: PyTorch + Transformers + PEFT
Distributed Training: torchrun with DDP

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "jiwon9703/KoLlama-3.1-8B-Instruct-qlora-sft-DDP-v0"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "한국의 수도는 어디인가요?"}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)

response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Evaluation

[TODO: Add evaluation results]

Limitations

This model inherits the limitations of the base Llama 3.1 model
May generate incorrect or biased information
Not suitable for critical applications without human oversight

Citation

If you use this model, please cite:

bibtex

@misc{kollama-3.1-8b-instruct-qlora,
  author = {jiwon9703},
  title = {KoLlama-3.1-8B-Instruct-qlora-sft-DDP-v0},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/jiwon9703/KoLlama-3.1-8B-Instruct-qlora-sft-DDP-v0}
}