Sea L L Ms v3 7 B Chat

Introduction

SeaLLMs-v3 is a series of large language models designed for Southeast Asian languages. The models achieve state-of-the-art performance in tasks such as world knowledge, mathematical reasoning, translation, and instruction following. They are fine-tuned to be more trustworthy, reducing hallucinations and providing culturally relevant responses.

Architecture

SeaLLMs supports multiple languages, including English, Chinese, Indonesian, Vietnamese, Thai, Tagalog, Malay, Burmese, Khmer, Lao, Tamil, and Javanese. The models excel in multilingual world knowledge and instruction-following capabilities.

Training

The models are evaluated on dimensions such as exam questions, instruction-following, mathematics, and translation. They have been trained to reduce hallucinations and ensure safety, particularly in Southeast Asian contexts.

Guide: Running Locally

1. Install Libraries: Make sure you have the latest version of transformers (>4.40) and, optionally, vllm for fast inference.
2. Set Up Environment: Use a cloud GPU service like AWS, Google Cloud, or Azure to handle the models efficiently.
3. Load Model and Tokenizer:

   from transformers import AutoModelForCausalLM, AutoTokenizer
   
   device = "cuda"
   model = AutoModelForCausalLM.from_pretrained(
     "SeaLLMs/SeaLLMs-v3-7B-Chat",
     torch_dtype=torch.bfloat16, 
     device_map=device
   )
   tokenizer = AutoTokenizer.from_pretrained("SeaLLMs/SeaLLMs-v3-7B-Chat")
   

4. Conduct Inference: Use the tokenizer and model to generate responses from user prompts, as demonstrated in the code snippets provided.
5. Use TextStreamer for Conversations: Implement a loop to interact continuously with the model.

License

The SeaLLMs models are released under the terms specified in the SeaLLMs Terms of Use. The license details are available here. Users must comply with these terms and are responsible for ensuring that deployments adhere to local regulations and governance.