elinas/Llama-3-13B-Instruct

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:15BQuant:FP8Context Size:8kTool Calling:SupportedPublished:May 7, 2024License:llama3Architecture:Transformer Featherless Exclusive Cold

elinas/Llama-3-13B-Instruct is a 15 billion parameter language model, merged from Meta-Llama-3-8B-Instruct using the passthrough method. This model aims to provide a mid-sized Llama 3 variant, suitable as a base model for further fine-tuning. It is usable for chat and storywriting, leveraging the Llama 3 Instruct template, though it may exhibit some grammatical quirks. Its primary strength lies in creative writing tasks, while logical abilities like programming and math are noted to be deteriorated.

Loading preview...

elinas/Llama-3-13B-Instruct Overview

This model is a 15 billion parameter language model, created by elinas through a merge of the Meta-Llama-3-8B-Instruct model using the passthrough method via mergekit. The intention behind this merge was to develop a Llama 3 model in a "mid" size category, suitable primarily as a base for subsequent fine-tuning efforts.

Key Capabilities

  • Chat and Storywriting: The model is functional for general chat and creative story generation, utilizing the Llama 3 Instruct template.
  • Base Model: It is designed to serve as a foundational model for further fine-tuning to enhance specific capabilities.

Limitations and Recommendations

  • Logical Abilities: The merge process has reportedly deteriorated the model's performance in logical tasks such as programming, mathematics, and science.
  • Repetition Penalty: It is recommended to use no repetition penalty or a value less than 1.05 to avoid erratic output.
  • Grammatical Quirks: Users may encounter occasional grammatical inconsistencies.

Further Development

A fine-tuned version of this model, which reportedly offers improved performance, is available at elinas/Llama-3-13B-Instruct-ft.

Finetunes
4 models