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

Hugging Face
TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:8kPublished:Aug 18, 2024License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Warm

zake7749/gemma-2-2b-it-chinese-kyara-dpo is a 2.6 billion parameter Gemma-2-2b-it model fine-tuned by zake7749 using the Kyara (Knowledge Yielding Adaptive Retrieval Augmentation) method. This model is specifically optimized for Traditional Chinese language understanding and generation, addressing the scarcity of Traditional Chinese data in LLM training. It demonstrates improved performance over the base Gemma-2-2b-it model across various benchmarks, particularly in Chinese language evaluations, and is noted for its experimental Retrieval Augmented Generation (RAG) capabilities.

Loading preview...

Model Overview

This model, zake7749/gemma-2-2b-it-chinese-kyara-dpo, is a 2.6 billion parameter variant of Gemma-2-2b-it that has undergone full-parameter fine-tuning using the Kyara (Knowledge Yielding Adaptive Retrieval Augmentation) experimental project. Kyara aims to enhance language models' ability to adapt knowledge and improve comprehension, especially in Traditional Chinese, a language with relatively scarce training data.

Key Capabilities & Differentiators

  • Enhanced Traditional Chinese Performance: Outperforms the original Gemma-2-2b-it across various benchmarks, with significant improvements in Chinese language evaluations like TMMLUPlus and AlignBench (Traditional Chinese fold).
  • Knowledge Yielding Adaptive Retrieval Augmentation: Utilizes a novel method for dataset construction, including knowledge injection with retrieval augmentation and a focus on high-quality, diverse datasets.
  • Preference Learning (DPO): Incorporates Direct Preference Optimization (DPO) using both English and custom-built Chinese datasets to align responses better with human preferences and enhance mathematical/programming abilities.
  • Experimental RAG Feature: Benefits from RAG-related content during SFT, allowing for structured responses that appropriately cite reference documents.
  • Leading 2B-scale Model: As of its release, Kyara-2b-it is positioned as a leading competitor among 2B-scale models on the Open-LLM Leaderboard.

Limitations

  • Like most LLMs, Kyara may still exhibit hallucinations.
  • It has a tendency to quote references frequently when answering questions, which might sometimes lead to incorrect attributions if the underlying knowledge is flawed.

Finetunes
4 models