yanolja/YanoljaNEXT-EEVE-Instruct-10.8B

TEXT GENERATIONConcurrency Cost:1Model Size:15BQuant:FP8Ctx Length:8kTool Calling:SupportedPublished:Feb 22, 2024License:apache-2.0Architecture:Transformer0.2K Open Weights Cold

YanoljaNEXT-EEVE-Instruct-10.8B is a 10.8 billion parameter instruction-tuned causal language model developed by Yanolja. It is a Korean vocabulary-extended version of upstage/SOLAR-10.7B-v1.0, fine-tuned using Direct Preference Optimization (DPO) with Axolotl. This model is specifically optimized for conversational tasks in Korean, leveraging translated datasets like SlimOrca-Dedup and Ultrafeedback-Binarized-Preferences-Cleaned. It achieves an average score of 66.48 on the Open LLM Leaderboard, demonstrating strong performance across various benchmarks including MMLU and HellaSwag.

Loading preview...

Overview

YanoljaNEXT-EEVE-Instruct-10.8B is a 10.8 billion parameter instruction-tuned language model developed by Yanolja. It is built upon the yanolja/EEVE-Korean-10.8B-v1.0 base model, which itself is a Korean vocabulary-extended version of upstage/SOLAR-10.7B-v1.0. The model was fine-tuned using Direct Preference Optimization (DPO) with the Axolotl framework, focusing on enhancing its conversational abilities.

Key Capabilities

  • Korean Language Proficiency: Specifically designed and optimized for understanding and generating Korean text, leveraging a vocabulary expansion technique detailed in their technical report.
  • Instruction Following: Fine-tuned with instruction datasets to provide helpful, detailed, and polite answers in a chat-based format.
  • Preference Alignment: Utilizes Direct Preference Optimization (DPO) to align model outputs with human preferences, leading to more desirable responses.
  • Benchmark Performance: Achieves an average score of 66.48 on the Open LLM Leaderboard, with notable scores such as 64.23 on MMLU (5-Shot) and 83.04 on HellaSwag (10-Shot).

Training Data

The model was trained on Korean-translated versions of Open-Orca/SlimOrca-Dedup and argilla/ultrafeedback-binarized-preferences-preferences-cleaned datasets.

Good For

  • Developing Korean-language chatbots and conversational AI applications.
  • Tasks requiring instruction-following and polite, detailed responses in Korean.
  • Research and development in multilingual LLMs, particularly for Korean language integration.