aixsatoshi/Llama-3-8b-Cosmopedia-japanese

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:May 1, 2024License:llama3Architecture:Transformer0.0K Featherless Exclusive Cold

The aixsatoshi/Llama-3-8b-Cosmopedia-japanese model is an 8 billion parameter Llama-3 variant adapted for the Japanese language domain. It addresses Llama-3-8b's strong English output bias by leveraging a Japanese-translated version of the high-quality, Mixtral-generated Cosmopedia synthetic dataset. This model aims to transfer Llama-3-8b's advanced logical reasoning capabilities to Japanese contexts without degradation, making it suitable for Japanese-centric applications requiring strong reasoning.

Loading preview...

Overview

The aixsatoshi/Llama-3-8b-Cosmopedia-japanese model is an 8 billion parameter language model based on the Llama-3-8b architecture, specifically adapted for the Japanese language. While Llama-3-8b demonstrates high capabilities, it exhibits a strong bias towards English responses and reduced performance in Japanese contexts.

Key Adaptations and Training

This model's primary goal is to maintain Llama-3-8b's advanced logical reasoning while shifting its output language bias towards Japanese. This adaptation was achieved through additional training using a Japanese version of the Cosmopedia dataset.

  • Cosmopedia Dataset: Originally composed of high-quality, noise-free outputs from Mixtral 8x7B, focusing on core reasoning capabilities.
  • Japanese Translation: Since Cosmopedia is English-based and Mixtral struggles with Japanese expressions, the dataset was translated into Japanese using an external translation system. This resulted in datasets like aixsatoshi/cosmopedia-japanese-100k and aixsatoshi/cosmopedia-japanese-20k.
  • Target: The additional training on this Japanese-translated data aims to seamlessly integrate Llama-3-8b's logical reasoning into Japanese contexts and re-orient its output language preference.

Good for

  • Applications requiring strong logical reasoning in Japanese.
  • Use cases where Llama-3-8b's core capabilities are desired but with a Japanese output bias.
  • Developers looking for a Japanese-adapted Llama-3 variant for various NLP tasks.