--- language: - id - en - jv - su license: llama2 library_name: transformers tags: - komodo --- # Model Card for Komodo-7B-Base Komodo-7B-Base is a large language model that is developed through incremental pretraining and vocabulary expansion on top of Llama-2-7B-Base. This model can handle Indonesian, English and 11 regional languages of Indonesia. **Disclaimer** : This is not an instruction-tuned model, further fine-tuning is needed for downstream tasks. For example, people usually utilize the [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca) dataset for further fine-tuning on top of Llama-2-7B-Base model. Hence, there is no prompt template for this model. ## Model Details

### Model Description More details can be found in our paper: https://arxiv.org/abs/2403.09362 - **Developed by:** [Yellow.ai](https://yellow.ai/) - **Model type:** Decoder - **Languages:** English, Indonesian, Acehnese, Balinese, Banjarese, Buginese, Madurese, Minangkabau, Javanese, Dayak Ngaju, Sundanese, Toba Batak, Lampungnese - **License:** llama2 ## Usage Example Since this is a gated model, you need to logged in to your HF account before using the model. Below is one way to do this. You can get the HF Token from your profile (Profile -> Settings -> Access Tokens) ``` import huggingface_hub huggingface_hub.login("YOUR_HF_TOKEN") ``` Once you are logged in, you can start download and load the model & tokenizer. We wrote a custom decoding function for Komodo-7B, that's why we need to pass the `trust_remote_code=True`. The code also works without this parameter, but decoding process will not work as expected. ``` import torch from transformers import AutoTokenizer, AutoModelForCausalLM device = "cuda:0" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained("Yellow-AI-NLP/komodo-7b-base",trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained("Yellow-AI-NLP/komodo-7b-base",trust_remote_code=True) model = model.to(device) ``` Then, you can try using the model. ``` full_prompt = "Candi borobudur adalah" tokens = tokenizer(full_prompt, return_tensors="pt").to(device) output = model.generate(tokens["input_ids"], eos_token_id=tokenizer.eos_token_id) print(tokenizer.decode(output[0], skip_special_tokens=True)) # Candi borobudur adalah candi yang terletak di Magelang, Jawa Tengah. ``` ## Technical Specifications ### Model Architecture and Objective Komodo-7B is a decoder model using the Llama-2 architecture. | Parameter | Komodo-7B | |-----------------|:-----------:| | Layers | 32 | | d_model | 4096 | | head_dim | 32 | | Vocabulary | 35008 | | Sequence Length | 4096 | ### Tokenizer Details Recognizing the importance of linguistic diversity, we focused on enhancing our language model's proficiency in both Indonesian and regional languages. To achieve this, we systematically expanded the tokenizer's vocabulary by identifying and incorporating approximately 2,000 frequently used words specific to Indonesian and 1,000 words for regional languages that were absent in the Llama-2 model. The standard method for enhancing a vocabulary typically involves developing a new tokenizer and integrating it with the existing one. This technique has shown impressive results in projects like Chinese-LLaMA and Open-Hathi. The effectiveness of this strategy can be attributed to the significant linguistic distinctions between languages such as Chinese and Hindi when compared to English. In contrast, the Indonesian language employs the same Latin script as English, which presents a different set of challenges. We tested the traditional method, as well as a new approach where we included the top n words (not tokens) from the Indonesian vocabulary. We discovered that with the new approach, we could achieve better fertility scores by adding around 3000 new vocabulary words. Adding more than 3000 words did not significantly improve the fertility score further, but it increased the size of the embedding matrix, leading to longer training times. More details can be found in our paper: https://arxiv.org/abs/2403.09362 ### Training Data More details can be found in our paper: https://arxiv.org/abs/2403.09362 ### Training Procedure More details can be found in our paper: https://arxiv.org/abs/2403.09362 #### Preprocessing More details can be found in our paper: https://arxiv.org/abs/2403.09362 ## Evaluation & Results Please note that the benchmarking values below are based on our SFT Model, Komodo-7B-Instruct, while here we only release the base model, Komodo-7B-base. | Organization | Model Name | Indo MMLU | ID-EN | XCOPA-ID | Intent Classification | Colloquial Detection | NusaX-Senti | ID-Hate Speech | TydiQA-ID | Indosum | Average | |--------------|--------------------|-----------|-------|----------|-----------------------|----------------------|-------------|----------------|-----------|---------|---------| | OpenAI | GPT-3.5-turbo-0301 | 51.3 | 64.5 | 70.0 | 82.0 | 64.1 | 47.2 | 68.0 | 85.3 | 41.0 | 63.7 | | OpenAI | GPT-3.5-turbo-0613 | 52.7 | 66.8 | 88.2 | 84.0 | 75.1 | 63.3 | 63.7 | 86.4 | 40.0 | 68.9 | | OpenAI | GPT-3.5-turbo-1106 | 53.3 | 69.7 | 89.3 | 84.0 | 64.2 | 59.8 | 56.6 | 88.0 | 42.0 | 67.4 | | OpenAI | GPT-4-preview-1106 | 69.8 | 78.0 | 98.3 | 89.0 | 92.7 | 66.1 | 73.4 | 72.0 | 33.0 | 74.7 | | Meta | Llama-2-7B-Chat | 30.4 | 45.6 | 41.5 | 57.0 | 31.4 | 2.9 | 41.3 | 11.7 | 34.0 | 32.9 | | Meta | Llama-2-13B-Chat | 32.0 | 61.7 | 38.0 | 59.0 | 31.1 | 58.7 | 57.2 | 71.9 | 40.0 | 50.0 | | Google | Gemma-7B-it | 37.4 | 73.6 | 57.7 | 77.1 | 18.8 | 44.2 | 54.8 | 73.3 | 44.0 | 53.4 | | Mistral | Mixtral-8x7B-v0.1-Instruct | 45.2 | 57.8 | 88.7 | 86.0 | 41.1 | 52.8 | 68.8 | 90.3 | 14.0 | 60.5 | | AISingapore | Sealion-7B-Instruct-NC | 23.9 | 26.9 | 41.3 | 37.0 | 41.8 | 30.7 | 57.3 | 65.3 | 26.0 | 38.9 | | Cohere | Aya-101-13B | 47.7 | 47.3 | 84.0 | 64.0 | 18.9 | 74.6 | 72.7 | 81.3 | 39.0 | 58.8 | | MBZUAI | Bactrian-X-Llama-7B | 23.6 | 43.2 | 45.3 | 42.0 | 50.3 | 44.5 | 42.4 | 65.0 | 15.0 | 41.3 | | Alibaba | Qwen-1.5-7B-chat | 40.0 | 56.0 | 29.5 | 85.0 | 41.8 | 58.7 | 63.9 | 51.22 | 29.0 | 50.6 | | Yellow.ai | Komodo-7B-Instruct | 43.2 | 90.5 | 79.6 | 84.0 | 73.6 | 79.3 | 56.2 | 90.3 | 43.0 | 71.1 |

More details can be found in our paper: https://arxiv.org/abs/2403.09362 ### Infrastructure | Training Details | Komodo-7B | |----------------------|:------------:| | AWS EC2 p4d.24xlarge | 1 instances | | Nvidia A100 40GB GPU | 8 | | Training Duration | 300 hours | ## Citation ``` @misc{owen2024komodo, title={Komodo: A Linguistic Expedition into Indonesia's Regional Languages}, author={Louis Owen and Vishesh Tripathi and Abhay Kumar and Biddwan Ahmed}, year={2024}, eprint={2403.09362}, archivePrefix={arXiv}, primaryClass={cs.CL} } ``` ## Model Card Authors [Louis Owen](https://www.linkedin.com/in/louisowen/)
[Vishesh Tripathi](https://www.linkedin.com/in/vishesh-tripathi/)
[Abhay Kumar](https://www.linkedin.com/in/akanyaani/)
[Biddwan Ahmed](https://www.linkedin.com/in/biddwan-ahmed-917333126/)