vitruv/vitruv_2

TEXT GENERATIONConcurrency Cost:1Model Size:15BQuant:FP8Ctx Length:8kTool Calling:SupportedPublished:Mar 20, 2024License:apache-2.0Architecture:Transformer Open Weights Cold

vitruv/vitruv_2 is a 15 billion parameter language model developed by Virtruv, building upon vitruv/vitruv1. This model is specifically fine-tuned with a focus on mathematical tasks in Korean, utilizing a combination of translated mathematical datasets and Korean cultural and conversational data. It is designed to enhance performance in Korean-language mathematical reasoning and related applications.

Loading preview...

vitruv/vitruv_2: Korean Math-Focused LLM

vitruv/vitruv_2 is a 15 billion parameter large language model developed by Virtruv, evolving from its predecessor, vitruv/vitruv1. The primary focus of this iteration is to enhance performance in Korean-language mathematical tasks.

Key Capabilities & Training

This model has been extensively trained on a diverse set of Korean and translated datasets, specifically curated to improve its mathematical reasoning and general Korean understanding:

  • Mathematical Focus: Significant effort was placed on incorporating mathematical data, including a sampled and translated version of the GAIR/MathPile dataset, alongside the kyujinpy/KOR-gugugu-platypus-set.
  • Korean Language Proficiency: Training also included general Korean datasets such as traintogpb/aihub-koen-translation-integrated-tiny-100k.
  • Cultural & Conversational Context: To broaden its understanding of Korean nuances, the model was further trained on AIHUB datasets comprising Korean cultural content (movie/drama scripts) and professional telephone consultation records.

Good For

  • Applications requiring mathematical problem-solving in Korean.
  • Tasks benefiting from a model with a strong foundation in general Korean language understanding.
  • Use cases that could leverage its exposure to Korean cultural and conversational data.