Aya Collection
CohereMultilingual instruction-tuning dataset spanning more than 114 languages, built from templated and translated subsets for tasks including classification, summarization and translation. It holds hundreds of millions of examples across 64 subsets, and was created by Cohere Labs as part of the Aya open-science initiative.
Openness
5 high confidence- license
- apache-2.0
- gated
- false
- card
- present
Ungated, Apache-2.0 licensed, with a complete dataset card.
- https://huggingface.co/api/datasets/CohereLabs/aya_collection recorded 2026-09-29
Hub API record for the dataset whose JSON carries a gated field set to false (public, ungated) and the tags "license:apache-2.0" and "task_categories:text-classification"
- https://huggingface.co/datasets/CohereLabs/aya_collection/raw/main/README.md recorded 2026-09-29
Raw dataset card front matter and licensing section, stating "license: apache-2.0" and "This dataset can be used for any purpose, whether academic or commercial, under the terms of the" Apache 2.0 License, and a summary of "513 million instances of prompts and completions covering a wide range of tasks"
Adoption
3 high confidenceAdoption is measured by Hugging Face downloads of CohereLabs/aya_collection over the trailing month, the only usage channel on record for Aya Collection.
- https://huggingface.co/api/datasets/CohereLabs/aya_collection recorded 2026-08-13
11,190 downloads in the trailing 30 days for CohereLabs/aya_collection
Capability
4 high confidenceAya 101 and Aya 23 are both documented as fine-tuned on the Aya Collection, and the Aya 23 paper reports it outperforming Gemma, Mistral and Mixtral on a wide range of tasks. The Aya Expanse paper reports win rates of up to 76.6% for its 8B and 32B models against Gemma 2, Qwen 2.5 and Llama 3.1, but no primary Aya Expanse source names this collection as its training data.
- https://arxiv.org/abs/2402.06619 recorded 2026-09-29
Aya Collection paper abstract stating "we create the most extensive multilingual collection to date, comprising 513 million instances through templating and translating existing datasets across 114 languages"
- https://arxiv.org/abs/2405.15032 recorded 2026-09-29
Aya 23 report abstract stating it focuses on "pairing a highly performant pre-trained model with the recently released Aya collection" and that "Aya 23 outperforms both previous massively multilingual models like Aya 101 for the languages it covers, as well as widely used models like Gemma, Mistral and Mixtral"
- https://arxiv.org/abs/2412.04261 recorded 2026-09-29
Aya Expanse paper abstract stating "Aya Expanse 8B and 32B outperform leading open-weight models in their respective parameter classes, including Gemma 2, Qwen 2.5, and Llama 3.1, achieving up to a 76.6% win-rate"; searched the body for Aya Collection and aya_collection: none
- https://huggingface.co/blog/aya-expanse recorded 2026-09-29
Aya Expanse blog reporting Aya Expanse 8B with "win rates ranging from 60.4% to 70.6%" and mentioning "Aya 101" accompanied by the Aya Collection, without saying the collection trained Aya Expanse
Verified 2026-09-27