KaraKaraWarehouse/L3.1-70b-Ginny

TEXT GENERATIONPricing:Input $2.88 / Output $2.88Concurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2024Architecture:Transformer Featherless Exclusive Cold

L3.1-70b-Ginny is a 70 billion parameter language model developed by KaraKaraWarehouse, built by merging several Llama 3.1-based models including NousResearch/Hermes-3-Llama-3.1-70B and cyberagent/Llama-3.1-70B-Japanese-Instruct-2407. This model is specifically designed to enhance Japanese language capabilities and general instruction following, leveraging a 32K token context length. It aims to provide improved performance in Japanese instruction and potential translation tasks through its unique merge composition.

Loading preview...

L3.1-70b-Ginny: A Merged Llama 3.1 Model

L3.1-70b-Ginny is a 70 billion parameter language model created by KaraKaraWarehouse through a merge of four distinct Llama 3.1-based models using LazyMergekit. The base model for this merge is NousResearch/Hermes-3-Llama-3.1-70B, combined with Fizzarolli/L3.1-70b-glitz-v0.2, cyberagent/Llama-3.1-70B-Japanese-Instruct-2407, and Sao10K/L3-70B-Euryale-v2.1.

Key Capabilities & Design Choices

  • Enhanced Japanese Language Support: A primary goal of this merge was to boost Japanese capabilities by incorporating cyberagent's Llama-3.1-70B-Japanese-Instruct-2407. Initial testing indicates success, with the model showing improved performance in Japanese and potential for translation tasks.
  • Instruction Following: The model is designed for general instruction following, inheriting capabilities from its constituent instruction-tuned models.
  • Context Length: It supports a context length of 32,768 tokens.

Usage Considerations

  • Chat Format: The model utilizes the Llama 3.1 Instruct chat format.
  • Output Style: Users may find the model can be quite verbose; lowering the temperature during generation is recommended to mitigate rambly outputs.

When to Use This Model

This model is particularly suitable for applications requiring strong Japanese language understanding, generation, or translation, alongside general instruction-following tasks. Its merged architecture aims to combine the strengths of its components for a versatile 70B parameter model.

Finetunes
1 models