--- library_name: transformers license: cc-by-nc-4.0 language: - en - ko base_model: - nlpai-lab/KULLM3 - upstage/SOLAR-10.7B-v1.0 --- # This repository is the awq quantization version of KULLM3. The quantization was carried out in a custom branch of [autoawq](https://github.com/casper-hansen/AutoAWQ/). The hyperparameters for quantization are as follows. ```{ "zero_point": True, "q_group_size": 128, "w_bit": 4, "version": "GEMM" }``` It worked using [vllm](https://github.com/vllm-project/vllm). It may not work with other frameworks as they have not been tested. # KULLM3 Introducing KULLM3, a model with advanced instruction-following and fluent chat abilities. It has shown remarkable performance in instruction-following, speficially by closely following gpt-3.5-turbo. To our knowledge, It is one of the best publicly opened Korean-speaking language models. For details, visit the [KULLM repository](https://github.com/nlpai-lab/KULLM) ### Model Description This is the model card of a πŸ€— transformers model that has been pushed on the Hub. - **Developed by:** [NLP&AI Lab](http://nlp.korea.ac.kr/) - **Language(s) (NLP):** Korean, English - **License:** CC-BY-NC 4.0 - **Finetuned from model:** [upstage/SOLAR-10.7B-Instruct-v1.0](https://huggingface.co/upstage/SOLAR-10.7B-Instruct-v1.0) ## Example code ### Install Dependencies ```bash pip install torch transformers==4.38.2 accelerate ``` - In transformers>=4.39.0, generate() does not work well. (as of 2024.4.4.) ### Python code ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer MODEL_DIR = "nlpai-lab/KULLM3" model = AutoModelForCausalLM.from_pretrained(MODEL_DIR, torch_dtype=torch.float16).to("cuda") tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR) streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) s = "κ³ λ €λŒ€ν•™κ΅μ— λŒ€ν•΄μ„œ μ•Œκ³  μžˆλ‹ˆ?" conversation = [{'role': 'user', 'content': s}] inputs = tokenizer.apply_chat_template( conversation, tokenize=True, add_generation_prompt=True, return_tensors='pt').to("cuda") _ = model.generate(inputs, streamer=streamer, max_new_tokens=1024) # λ„€, κ³ λ €λŒ€ν•™κ΅μ— λŒ€ν•΄ μ•Œκ³  μžˆμŠ΅λ‹ˆλ‹€. κ³ λ €λŒ€ν•™κ΅λŠ” λŒ€ν•œλ―Όκ΅­ μ„œμšΈμ— μœ„μΉ˜ν•œ 사립 λŒ€ν•™κ΅λ‘œ, 1905년에 μ„€λ¦½λ˜μ—ˆμŠ΅λ‹ˆλ‹€. 이 λŒ€ν•™κ΅λŠ” ν•œκ΅­μ—μ„œ κ°€μž₯ 였래된 λŒ€ν•™ 쀑 ν•˜λ‚˜λ‘œ, λ‹€μ–‘ν•œ ν•™λΆ€ 및 λŒ€ν•™μ› ν”„λ‘œκ·Έλž¨μ„ μ œκ³΅ν•©λ‹ˆλ‹€. κ³ λ €λŒ€ν•™κ΅λŠ” 특히 법학, κ²½μ œν•™, μ •μΉ˜ν•™, μ‚¬νšŒν•™, λ¬Έν•™, κ³Όν•™ λΆ„μ•Όμ—μ„œ 높은 λͺ…성을 가지고 μžˆμŠ΅λ‹ˆλ‹€. λ˜ν•œ, 슀포츠 λΆ„μ•Όμ—μ„œλ„ ν™œλ°œν•œ ν™œλ™μ„ 보이며, λŒ€ν•œλ―Όκ΅­ λŒ€ν•™ μŠ€ν¬μΈ μ—μ„œ μ€‘μš”ν•œ 역할을 ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. κ³ λ €λŒ€ν•™κ΅λŠ” ꡭ제적인 ꡐλ₯˜μ™€ ν˜‘λ ₯에도 적극적이며, μ „ 세계 λ‹€μ–‘ν•œ λŒ€ν•™κ³Όμ˜ ν˜‘λ ₯을 톡해 κΈ€λ‘œλ²Œ 경쟁λ ₯을 κ°•ν™”ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€. ``` ## Training Details ### Training Data - [vicgalle/alpaca-gpt4](https://huggingface.co/datasets/vicgalle/alpaca-gpt4) - Mixed Korean instruction data (gpt-generated, hand-crafted, etc) - About 66000+ examples used totally ### Training Procedure - Trained with fixed system prompt below. ```text 당신은 κ³ λ €λŒ€ν•™κ΅ NLP&AI μ—°κ΅¬μ‹€μ—μ„œ λ§Œλ“  AI μ±—λ΄‡μž…λ‹ˆλ‹€. λ‹Ήμ‹ μ˜ 이름은 'KULLM'으둜, ν•œκ΅­μ–΄λ‘œλŠ” 'ꡬ름'을 λœ»ν•©λ‹ˆλ‹€. 당신은 λΉ„λ„λ•μ μ΄κ±°λ‚˜, μ„±μ μ΄κ±°λ‚˜, λΆˆλ²•μ μ΄κ±°λ‚˜ λ˜λŠ” μ‚¬νšŒ ν†΅λ…μ μœΌλ‘œ ν—ˆμš©λ˜μ§€ μ•ŠλŠ” λ°œμ–Έμ€ ν•˜μ§€ μ•ŠμŠ΅λ‹ˆλ‹€. μ‚¬μš©μžμ™€ 즐겁게 λŒ€ν™”ν•˜λ©°, μ‚¬μš©μžμ˜ 응닡에 κ°€λŠ₯ν•œ μ •ν™•ν•˜κ³  μΉœμ ˆν•˜κ²Œ μ‘λ‹΅ν•¨μœΌλ‘œμ¨ μ΅œλŒ€ν•œ 도와주렀고 λ…Έλ ₯ν•©λ‹ˆλ‹€. 질문이 μ΄μƒν•˜λ‹€λ©΄, μ–΄λ–€ 뢀뢄이 μ΄μƒν•œμ§€ μ„€λͺ…ν•©λ‹ˆλ‹€. 거짓 정보λ₯Ό λ°œμ–Έν•˜μ§€ μ•Šλ„λ‘ μ£Όμ˜ν•©λ‹ˆλ‹€. ``` ## Evaluation - Evaluation details such as testing data, metrics are written in [github](https://github.com/nlpai-lab/KULLM). - Without system prompt used in training phase, KULLM would show lower performance than expect. ### Results ## Citation ```text @misc{kullm, author = {NLP & AI Lab and Human-Inspired AI research}, title = {KULLM: Korea University Large Language Model Project}, year = {2023}, publisher = {GitHub}, journal = {GitHub repository}, howpublished = {\url{https://github.com/nlpai-lab/kullm}}, } ```