zake7749/gemma-2-2b-it-chinese-kyara-dpo

Overview

Kyara (Knowledge Yielding Adaptive Retrieval Augmentation) is a 2.6 billion parameter model based on Gemma-2-2b-it, developed by zake7749. Its primary goal is to improve language models through adaptive knowledge retrieval, with a strong focus on enhancing comprehension in underrepresented languages like Traditional Chinese. The model was fine-tuned using a full-parameter approach and has shown significant improvements over the original Gemma-2-2b-it, particularly in Chinese language benchmarks.

Key Capabilities

• Enhanced Chinese Language Performance: Outperforms the base Gemma-2-2b-it on various Chinese language evaluations, including TMMLUPlus and Chinese-Reason-Bench.
• Knowledge Retrieval Integration: Trained with a unique method that incorporates Retrieval Augmented Generation (RAG) during the Supervised Fine-Tuning (SFT) phase, making it suitable for RAG-based applications.
• Preference Learning: Utilizes Direct Preference Optimization (DPO) with custom Chinese and English datasets to align responses with human preferences and improve mathematical reasoning and programming skills.
• Leading 2B-scale Model: As of its release, Kyara-2b-it is a leading competitor among 2B-scale models on the OpenLLM Leaderboard.

Good For

• Applications requiring strong performance in Traditional Chinese language understanding and generation.
• Use cases benefiting from Retrieval Augmented Generation (RAG), where the model can effectively integrate external knowledge.
• Tasks involving mathematical reasoning and programming, due to its preference learning strategy.

Limitations

• Like most LLMs, Kyara may still exhibit hallucinations. Users should exercise caution, especially when the model cites references, as it might occasionally misattribute sources.