Model Overview

ghost-7b-v0.9.0 is a 7 billion parameter language model developed by Lam H, building upon the HuggingFaceH4/zephyr-7b-beta architecture. Its primary distinction lies in its bilingual fine-tuning, utilizing a small synthetic dataset (approximately 200MB) split equally between English and Vietnamese content. This targeted training aims to enhance its proficiency in both languages for various applications.

Key Capabilities

• Bilingual Proficiency: Excels in both English and Vietnamese, demonstrated through examples of question answering and language identification in both languages.
• Conversational AI: Designed to support multi-turn chat interactions, allowing for dynamic and context-aware dialogues.
• General Task Completion: Capable of handling a variety of natural language processing tasks beyond just conversation.
• MIT License: Available under the permissive MIT license, allowing for broad use and modification.

Performance Insights

Evaluations on the Open LLM Leaderboard show an average score of 56.89, with specific scores including 77.93 on HellaSwag and 55.09 on MMLU. Crucially, the model has also been assessed using the VMLU evaluation suite, which measures performance on Vietnamese-specific tasks. While VMLU results were obtained with 4-bit quantization, they indicate the model's specialized capabilities in the Vietnamese language across various domains like STEM, humanities, and social sciences.

Good For

• Applications requiring robust bilingual support in English and Vietnamese.
• Developing chatbots and conversational agents for a Vietnamese-speaking audience or for dual-language interactions.
• Tasks that benefit from a model fine-tuned on a balanced dataset for these two specific languages.