yanolja/YanoljaNEXT-EEVE-Instruct-7B-v2-Preview

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 23, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

YanoljaNEXT-EEVE-Instruct-7B-v2-Preview is a 7.6 billion parameter instruction-following large language model, fine-tuned from Qwen2.5-7B by yanolja. It features expanded Korean vocabulary and a unique hybrid architecture that allows for optional step-by-step reasoning. This model is primarily designed for enhanced Korean language understanding, generation, and specific tasks like English-to-Korean translation, while also supporting general instruction following, math, and coding.

Loading preview...

YanoljaNEXT-EEVE-Instruct-7B-v2-Preview Overview

This model is a 7.6 billion parameter instruction-following large language model, developed by yanolja and derived from Qwen2.5-7B. It has been specifically enhanced for Korean language understanding and generation through significant vocabulary expansion, adding 6,257 Korean tokens. A key differentiator is its hybrid nature, allowing users to optionally activate a step-by-step reasoning process using <think> tags before the model provides its final answer, which is particularly useful for complex tasks like math, coding, and detailed translation.

Key Capabilities

  • Enhanced Korean Language: Optimized for Korean understanding and generation due to expanded vocabulary.
  • Step-by-Step Reasoning: Supports explicit reasoning processes for complex problem-solving.
  • Specialized Translation: Includes a detailed prompt structure for high-quality English-to-Korean translation.
  • General Instruction Following: Capable of handling a wide range of conversational and instructional prompts.

Training and Limitations

The model was fine-tuned using a combination of datasets, including distilled data from DeepSeek-R1, HuggingFaceTB/smoltalk, HuggingFaceH4/ultrafeedback_binarized, and the AI Hub Korean Conversation Summary dataset. As a "Preview" release, it may have unoptimized performance, bugs, and is subject to general LLM limitations such as factual inaccuracies (hallucinations) and potential biases. Quantitative evaluation results are pending.

Finetunes
1 models