Openthaigpt 1.0.0 70b Chat
Openthaigpt 1.0.0 70b Chat is a cutting-edge AI model that sets a new standard for Thai language processing. What makes it unique? It's the first 70-billion-parameter Thai open-source language model, achieving higher scores on Thai exams than OpenAI GPT 3.5, Google Gemini, and Claude 3 Haiku. But how does it do it? By incorporating over 10,000 frequently used Thai words into its dictionary, significantly boosting its response speed. This model is not just fast, it's also capable of understanding and processing input contexts of up to 4096 Thai words, allowing for detailed and complex instructions. Want to know more about its capabilities? It supports extended conversations across multiple turns and even the use case of Retrieval Augmented Generation for enriched response generation. So, what can you do with this model? From answering questions to generating text, it's designed to be a helpful assistant. With its advanced architecture and training on over 65 billion Thai language words, it's no wonder it achieves the highest average scores across several Thai language exams.
Table of Contents
Model Overview
The OpenThaiGPT 70b model is a cutting-edge Thai language chat model, boasting 70 billion parameters. It’s specifically fine-tuned for Thai instructions and has been enhanced with over 10,000 commonly used Thai words, resulting in a significant boost in response speed.
Capabilities
This model is designed to engage in extended conversations across multiple turns, making it suitable for chatbots and virtual assistants. It can generate high-quality text based on a given prompt or context. Additionally, it supports Retrieval Augmented Generation (RAG), which enables it to retrieve relevant information from a knowledge base and incorporate it into its responses.
Primary Tasks
- Conversational AI: Engage in extended conversations across multiple turns.
- Text Generation: Generate high-quality text based on a given prompt or context.
- Retrieval Augmented Generation (RAG): Retrieve relevant information from a knowledge base and incorporate it into responses.
Strengths
- Leading-edge Thai language understanding: Achieves the highest average scores across several Thai language exams, outperforming other open-source Thai large language models.
- Fast response generation: Increased response speed by tenfold, thanks to the addition of frequently used Thai words.
- Detailed and complex instructions: Can understand and process input contexts of up to
4096Thai words, allowing for detailed and complex instructions.
Performance
This model showcases remarkable performance with exceptional speed, accuracy, and efficiency in various tasks.
Speed
- Generation speeds increased by tenfold thanks to the addition of
10,000frequently used Thai words to the model’s dictionary. - Can process input contexts of up to
4096Thai words, allowing for detailed and complex instructions.
Accuracy
- Achieves the highest average scores across several Thai language exams compared to all other open-source Thai LLMs.
- Outperforms ==OpenAI GPT 3.5==, ==Google Gemini==, and ==Claude 3 Haiku== in Thai exams.
Efficiency
- Pretrained on a foundation of over
65 billionThai language words and fine-tuned with over1 millionThai instruction examples. - Supports extended conversations across multiple turns and the use case of Retrieval Augmented Generation (RAG) for enriched response generation.
Limitations
While this model is a powerful tool, it’s not perfect. Here are some of its limitations:
Lack of Common Sense
- May lack common sense or real-world experience, leading to responses that are not practical or relevant in everyday situations.
Limited Domain Knowledge
- May not have the same level of expertise as a medical professional or a lawyer.
Biased Responses
- Can reflect biases present in the data it was trained on, resulting in responses that are discriminatory or unfair.
Overfitting
- May overfit to the training data, which means it may not generalize well to new, unseen data.
Dependence on Context
- Relies heavily on context to generate accurate responses. If the context is unclear or incomplete, the model may struggle to provide a relevant answer.
Limited Ability to Reason
- May not be able to follow complex logical arguments or understand subtle nuances in language.
Vulnerability to Adversarial Attacks
- Can be vulnerable to adversarial attacks, which are designed to manipulate the model’s responses.
Limited Support for Multimodal Input
- Primarily designed for text-based input and output. May not be able to handle multimodal input, such as images or audio, as effectively.
Dependence on Hardware
- Requires significant computational resources to run efficiently. This can be a challenge for devices with limited hardware capabilities.
Format
This model is a large language model based on the LLaMA v2 architecture, specifically designed for the Thai language. It has been fine-tuned for Thai instructions and enhanced with over 10,000 commonly used Thai words in its dictionary.
Architecture
- Has a transformer-based architecture, which allows it to process input sequences of up to
4096Thai words.
Data Formats
- Accepts input in the form of tokenized text sequences, with a specific prompt format.
Prompt Format
- The prompt format is based on Llama2 with a small modification, adding
###to specify the context part. The format is as follows:
\<s>[INST]
{system_prompt}
{human_turn1}####{context_turn1} [/INST]{assistant_turn1}
Input and Output
- Can handle single-turn and multi-turn conversations, as well as conversations with context (RAG). The input prompt should be in the specified format, and the output will be a generated response.
Examples
- Single Turn Conversation Example:
\<s>[INST]
You are a question answering assistant. Answer the question as truthful and helpful as possible คุณคือผู้ช่วยตอบคำถาม จงตอบคำถามอย่างถูกต้องและมีประโยชน์ที่สุด
สวัสดีครับ [/INST]
- Multi Turn Conversation Example:
\<s>[INST]
You are a question answering assistant. Answer the question as truthful and helpful as possible คุณคือผู้ช่วยตอบคำถาม จงตอบคำถามอย่างถูกต้องและมีประโยชน์ที่สุด
สวัสดีครับ [/INST]
สวัสดีค่ะ มีคำถามอะไร ถามได้เลย
...
- Multi Turn Conversation with Context (RAG) Example:
\<s>[INST]
You are a question answering assistant. Answer the question as truthful and helpful as possible คุณคือผู้ช่วยตอบคำถาม จงตอบคำถามอย่างถูกต้องและมีประโยชน์ที่สุด
กรุงเทพมีพื้นที่เท่าไร่###กรุงเทพมหานคร เป็นเมืองหลวง นครและมหานครที่มีประชากรมากที่สุดของประเทศไทย กรุงเทพมหานครมีพื้นที่ทั้งหมด 1,568.737 ตร.กม. มีประชากรตามทะเบียนราษฎรกว่า 8 ล้านคน [/INST]
...
GPU Memory Requirements
- Requires a significant amount of GPU memory, with the following requirements:
| Number of Parameters | FP 16 bits | 8 bits (Quantized) | 4 bits (Quantized) |
|---|---|---|---|
| 7b | 24 GB | 12 GB | 6 GB |
| 13b | 48 GB | 24 GB | 12 GB |
| 70b | 192 GB | 96 GB | 48 GB |
