--- language: - ko - en license: llama3 library_name: transformers tags: - llama - llama-3 base_model: - meta-llama/Meta-Llama-3-8B-Instruct datasets: - MarkrAI/KoCommercial-Dataset --- # Waktaverse-Llama-3-KO-8B-Instruct Model Card ## Model Details ![image/webp](https://cdn-uploads.huggingface.co/production/uploads/65d6e0640ff5bc0c9b69ddab/Va78DaYtPJU6xr4F6Ca4M.webp) Waktaverse-Llama-3-KO-8B-Instruct is a Korean language model developed by Waktaverse AI team. This large language model is a specialized version of the Meta-Llama-3-8B-Instruct, tailored for Korean natural language processing tasks. It is designed to handle a variety of complex instructions and generate coherent, contextually appropriate responses. - **Developed by:** Waktaverse AI - **Model type:** Large Language Model - **Language(s) (NLP):** Korean, English - **License:** [Llama3](https://llama.meta.com/llama3/license) - **Finetuned from model:** [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) ## Model Sources - **Repository:** [GitHub](https://github.com/PathFinderKR/Waktaverse-LLM/tree/main) - **Paper :** [More Information Needed] ## Uses ### Direct Use The model can be utilized directly for tasks such as text completion, summarization, and question answering without any fine-tuning. ### Out-of-Scope Use This model is not intended for use in scenarios that involve high-stakes decision-making including medical, legal, or safety-critical areas due to the potential risks of relying on automated decision-making. Moreover, any attempt to deploy the model in a manner that infringes upon privacy rights or facilitates biased decision-making is strongly discouraged. ## Bias, Risks, and Limitations While Waktaverse Llama 3 is a robust model, it shares common limitations associated with machine learning models including potential biases in training data, vulnerability to adversarial attacks, and unpredictable behavior under edge cases. There is also a risk of cultural and contextual misunderstanding, particularly when the model is applied to languages and contexts it was not specifically trained on. ## How to Get Started with the Model You can run conversational inference using the Transformers Auto classes. We highly recommend that you add Korean system prompt for better output. Adjust the hyperparameters as you need. ### Example Usage ```python import torch from transformers import AutoTokenizer, AutoModelForCausalLM device = ( "cuda:0" if torch.cuda.is_available() else # Nvidia GPU "mps" if torch.backends.mps.is_available() else # Apple Silicon GPU "cpu" ) model_id = "PathFinderKR/Waktaverse-Llama-3-KO-8B-Instruct" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map=device, ) ################################################################################ # Generation parameters ################################################################################ num_return_sequences=1 max_new_tokens=1024 temperature=0.6 top_p=0.9 repetition_penalty=1.1 def prompt_template(system, user): return ( "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n" f"{system}<|eot_id|>" "<|start_header_id|>user<|end_header_id|>\n\n" f"{user}<|eot_id|>" "<|start_header_id|>assistant<|end_header_id|>\n\n" ) def generate_response(system ,user): prompt = prompt_template(system, user) input_ids = tokenizer.encode( prompt, add_special_tokens=False, return_tensors="pt" ).to(device) outputs = model.generate( input_ids=input_ids, pad_token_id=tokenizer.eos_token_id, num_return_sequences=num_return_sequences, max_new_tokens=max_new_tokens, do_sample=True, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty ) return tokenizer.decode(outputs[0], skip_special_tokens=False) system_prompt = "다음 지시사항에 대한 응답을 작성해주세요." user_prompt = "피보나치 수열에 대해 설명해주세요." response = generate_response(system_prompt, user_prompt) print(response) ``` ### Example Output ```python <|begin_of_text|><|start_header_id|>system<|end_header_id|> 다음 지시사항에 대한 응답을 작성해주세요.<|eot_id|><|start_header_id|>user<|end_header_id|> 피보나치 수열에 대해 설명해주세요.<|eot_id|><|start_header_id|>assistant<|end_header_id|> 피보나치 수열은 0과 1 두 개의 숫자로 시작하는 무한정 길이가 있는 수열입니다. 이 수열의 각 요소를 다음 공식으로 계산합니다: F(n) = F(n-1) + F(n-2), 여기서 F(0) = 0, F(1) = 1입니다. 예를 들어, 첫 번째 10개의 피보나치 숫자는 다음과 같습니다: 0, 1, 1, 2, 3, 5, 8, 13, 21, 34 ``` ## Training Details ### Training Data The model is trained on the [MarkrAI/KoCommercial-Dataset](https://huggingface.co/datasets/MarkrAI/KoCommercial-Dataset), which consists of various commercial texts in Korean. ### Training Procedure The model training used LoRA for computational efficiency. 0.04 billion parameters(0.51% of total parameters) were trained. #### Training Hyperparameters ```python ################################################################################ # bitsandbytes parameters ################################################################################ load_in_4bit=True bnb_4bit_compute_dtype=torch.bfloat16 bnb_4bit_quant_type="nf4" bnb_4bit_use_double_quant=True ################################################################################ # LoRA parameters ################################################################################ task_type="CAUSAL_LM" target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] r=16 lora_alpha=32 lora_dropout=0.1 bias="none" ################################################################################ # TrainingArguments parameters ################################################################################ num_train_epochs=1 per_device_train_batch_size=1 gradient_accumulation_steps=1 gradient_checkpointing=True learning_rate=2e-5 lr_scheduler_type="cosine" warmup_ratio=0.1 optim = "paged_adamw_32bit" weight_decay=0.01 ################################################################################ # SFT parameters ################################################################################ max_seq_length=2048 packing=False ``` ## Evaluation ### Metrics - **Ko-HellaSwag:** - **Ko-MMLU:** - **Ko-Arc:** - **Ko-Truthful QA:** - **Ko-CommonGen V2:** ### Results
Benchmark Waktaverse Llama 3 8B Llama 3 8B
Ko-HellaSwag: 0 0
Ko-MMLU: 0 0
Ko-Arc: 0 0
Ko-Truthful QA: 0 0
Ko-CommonGen V2: 0 0
## Technical Specifications ### Compute Infrastructure #### Hardware - **GPU:** NVIDIA GeForce RTX 4080 SUPER #### Software - **Operating System:** Linux - **Deep Learning Framework:** Hugging Face Transformers, PyTorch ### Training Details - **Training time:** 18 hours - More details on [Weights & Biases](https://wandb.ai/pathfinderkr/Waktaverse-Llama-3-KO-8B-Instruct?nw=nwuserpathfinderkr) ## Citation **Waktaverse-Llama-3** ``` @article{waktaversellama3modelcard, title={Waktaverse Llama 3 Model Card}, author={AI@Waktaverse}, year={2024}, url = {https://huggingface.co/PathFinderKR/Waktaverse-Llama-3-KO-8B-Instruct} ``` **Llama-3** ``` @article{llama3modelcard, title={Llama 3 Model Card}, author={AI@Meta}, year={2024}, url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md} } ``` ## Model Card Authors [PathFinderKR](https://github.com/PathFinderKR)