What is EXAONE 3.5
EXAONE 3.5 is an open-source AI model launched by LG AI Research, including three versions with 2.4 billion, 7.8 billion, and 32 billion parameters. EXAONE 3.5 excels at long-text processing and performs excellently in benchmark tests, especially in real-world applications, long-text processing, and mathematics. The model uses retrieval-augmented generation technology and multi-step reasoning capabilities to effectively reduce misinformation and improve accuracy. LG plans to further expand its AI capabilities and launch the enterprise-level AI agent service ChatEXAONE, which features complex query analysis and user-defined search functions, equipped with encryption and privacy protection technology to ensure safe use within the company.
Main features of EXAONE 3.5
- Multi-version model support: Provides three models of different parameter scales to adapt to different application scenarios and computing resource constraints.
- Instruction-following capability: Has excellent instruction-following capability in real-world scenarios, achieving the highest scores in multiple benchmark tests.
- Long-context understanding: Performs excellently in long-text processing, effectively understanding and processing contexts up to 32K tokens.
- Bilingual capability: Excellent Korean and English bilingual capability, especially outstanding in Korean and English benchmark tests.
- Retrieval-augmented generation technology: Uses retrieval-augmented generation technology to generate answers based on reference documents or web search results.
- Multi-step reasoning capability: Has multi-step reasoning capability, effectively reducing the "hallucination" phenomenon and improving the accuracy of answers.
Technical principles of EXAONE 3.5
- Transformer architecture: Based on the latest decoder-only Transformer architecture, a deep learning model used for processing sequence data.
- Long-context processing: Uses long-context fine-tuning technology to extend the maximum context length from 4,096 tokens in EXAONE 3.0 to 32,768 tokens.
- Pre-training and post-training：
- Pre-training: The first stage uses a large training corpus for pre-training, and the second stage collects data and conducts pre-training for areas that need strengthening, especially enhancing long-context understanding capability.
- Post-training: Includes supervised fine-tuning (SFT) and preference optimization, strengthening the model's instruction-following capability and alignment with human preferences.
- Data compliance: Conducts AI compliance reviews during data collection, model training, and information provision to minimize legal risks.
- Retrieval-augmented generation (RAG) technology: Combines retrieval and generation, allowing the model to handle longer contexts and be applied in complex scenarios.
Project addresses of EXAONE 3.5
- Project official website：lgresearch.ai/blog/view
- GitHub repository：https://github.com/LG-AI-EXAONE/EXAONE-3.5
- HuggingFace model library：https://huggingface.co/collections/LGAI-EXAONE/exaone-35
- arXiv technical paper：https://arxiv.org/pdf/2412.04862
Application scenarios of EXAONE 3.5
- Chatbots and customer service: As the core of chatbots, handles customer queries and requests, providing 24*7 instant service.
- Language translation and cross-language understanding: Based on bilingual capability, assists with translation work and helps users from different language backgrounds communicate effectively.
- Content creation and editing: The model can generate creative copy, helping editors and writers expand ideas and improve the efficiency and quality of content creation.
- Education and Research: In the field of education, as an auxiliary tool, it helps students learn languages and answer academic questions.
- Information Retrieval and Knowledge Management: In enterprises, it helps employees quickly find the information they need, improving work efficiency and decision quality.