| --- |
| library_name: transformers |
| tags: |
| - llada |
| - sft |
| - instruction-tuning |
| - korean |
| - english |
| language: |
| - ko |
| - en |
| base_model: |
| - GSAI-ML/LLaDA-8B-Instruct |
| --- |
| |
| # Model Card for yevvonlim/Llada-8B-Instruct-Kor |
|
|
| <!-- Provide a quick summary of what the model is/does. --> |
|
|
| yevvonlim/Llada-8B-Instruct-Kor is an instruction-tuned variant of LLADA-8B designed for high-quality conversational responses in both Korean and English. Fine-tuned with supervised data, it excels at understanding and generating context-aware replies for chat applications. |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| This model is a supervised fine-tuned (SFT) version of [GSAI-ML/LLaDA-8B-Instruct], developed and shared by Sionic AI. It leverages parameter-efficient fine-tuning (PEFT) with LoRA to adapt the base LLADA-8B model to instruction-following tasks. |
|
|
| - **Developed by:** yevvonlim |
| - **Model type:** 8B-parameter encoder-only transformer |
| - **Language(s):** Korean, English |
| - **License:** Apache-2.0 |
| - **Fine-tuned from:** GSAI-ML/LLaDA-8B-Instruct |
|
|
| ### Model Sources |
|
|
| - **Repository:** https://github.com/yevvonlim/CAS4133-LLaDA-Kor |
| - **Paper:** N/A |
| - **Demo:** N/A |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| - Conversational agents and chatbots in Korean and English |
| - Instruction-following and question-answering tasks |
| - Assistive tools for writing, translation, and summarization |
|
|
| ### Out-of-Scope Use |
|
|
| - Tasks requiring specialized domain knowledge outside the training data |
| - Real-time high-stakes decision-making without human oversight |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - May produce incorrect or outdated facts |
| - Vulnerable to generating biased or stereotypical language present in training data |
|
|
| ### Recommendations |
|
|
| Users should review and verify model outputs before deployment in critical applications. Implement human-in-the-loop validation for high-stakes use cases. |
|
|
| ## How to Get Started with the Model |
|
|
| Use the code below to load and generate with the model. Ensure you have defined or imported the `generate_stream` function provided in the repository. |
|
|
| ```python |
| |
| from transformers import AutoTokenizer, AutoModel |
| |
| |
| device = "cuda" |
| model_path = "yevvonlim/Llada-8B-Instruct-Kor" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) |
| model = AutoModel.from_pretrained(model_path, trust_remote_code=True).to(device).eval() |
| |
| prompt = "6나누기 0은 뭐야? let's think step by step." |
| chat_input = tokenizer.apply_chat_template( |
| [{"role": "user", "content": prompt}], |
| add_generation_prompt=True, |
| tokenize=False, |
| ) |
| prompt_ids = tokenizer(chat_input, return_tensors="pt").input_ids.to(device) |
| |
| final_ids = model.generate(prompt_ids)[0, prompt_ids.shape[1]:] |
| print(tokenizer.decode(final_ids, skip_special_tokens=True)) |
| ``` |
|
|
| ## Contact |
|
|
| For issues or questions, please open an issue on the repo or contact ga06033@gmail.com |
