Aya Series by Cohere
The Aya Series is a family of open-source multilingual large language models developed by Cohere For AI. These models are designed to bridge language gaps globally and support research, covering up to 101 languages. The series includes several models tailored for specific use cases, such as multilingual text generation, instruction-following, and multimodal tasks.
Key Attributes
Current Models in the Aya Series:
Aya 101: Covers 101 languages with a focus on underserved languages.
Aya Expanse 32B: Optimized for depth and performance across 23 languages.
Aya Vision: A multimodal vision-language model supporting tasks like image captioning and OCR.
Multilingual Excellence:
Outperforms competitors like BLOOMZ and mT0 on multilingual benchmarks.
Open Source:
Released under the Apache 2.0 license with open weights for research and non-commercial use.
Scalable Model Sizes:
Available in parameter sizes ranging from 8B to 35B.
Innovative Training Techniques:
Synthetic data generation for rare languages.
Late-stage human feedback fine-tuning.
Example Use Cases
Multilingual translation and text generation.
Research applications in underserved languages.
Instruction-following tasks across diverse linguistic contexts.
Multimodal applications like image-to-text translation (Aya Vision).