Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning

Motivation

• Instruction fine-tuning datasets have focused almost entirely on English and a few data-rich languages, biasing model performance and reinforcing disparities for under-represented languages.These disparities can include language and dialect biases, security flaws, higher token costs, and latency for marginalized languages.Source: §1 · p. 3

Participatory data collection

• Aya’s year-long participatory initiative engaged 2,997 collaborators across 119 countries to produce 204,114 human-curated instruction-completion annotations in 65 languages.The annotations were collected from fluent speakers through the Aya Annotation Platform.Source: §1 · p. 4

Open resources

• Aya releases the Aya Annotation Platform code, Aya Dataset, Aya Collection, and Aya Evaluation Suite under a permissive Apache 2.0 License for researchers and practitioners.The datasets are openly available for download.Source: §1 · pp. 4–5

Aya Annotation Platform

• The Aya Annotation Platform supports multilingual instruction-style annotation across 182 languages, including dialects, and registered users from 119 countries and 134 languages.The platform was developed as a simple, intuitive interface for contributors to join the project.Source: §1 · pp. 4–5

Aya Collection

• The Aya Collection transforms 44 datasets into templated instruction-completion pairs and expands them through translation to 513M instances covering 114 languages.The collection includes tasks such as text classification, text generation, machine translation, paraphrasing, and open-domain question answering.Source: §1 · pp. 4–5

Aya Evaluation Suite

• The Aya Evaluation Suite provides multilingual open-ended generation evaluation with original, translated, and human-edited prompts spanning 7 directly curated languages and 101 translated languages.It includes 250 human-written prompts for each of 7 languages, 200 translated prompts for 101 languages, and human-edited prompts for 6 languages.Source: §1 · p. 5

2.1 Aya Annotation Platform

• The open-source Aya Annotation Platform supported data collection across 119 countries and 134 languages, prioritizing accessibility and ease of use for contributors unfamiliar with AI.The platform was designed to accommodate decentralized contributors worldwide and fully open-sourced its user-interface code.Source: §2.1 · pp. 5–6
• Aya supported both desktop and mobile browsers, with approximately 54% of users accessing it via desktop and 46% via mobile.Single Sign-On initially used Discord and later added Google after platform-specific access barriers were identified.Source: §2.1 · p. 6
• The platform prioritized 101 mT5-supported languages, but the final Aya Dataset covered 65 languages after contribution levels altered inclusion decisions.Wolof was added despite not being in the initial prioritized list because it received substantial contributions.Source: §2.1 · p. 6
• Aya included 2,997 registered users across 134 languages and 119 countries, with India contributing the most users at 346.More than two-thirds of users were aged 18–35; 68.1% identified as male, 28.5% as female, and 6.6% self-reported dialects.Source: Figure 2 · p. 6, §2.1 · p. 7

Geographic-Based Contribution Assessment

• 58.8% of Aya contributions came from Asia, 27.4% from Africa, and 13.8% from Europe, Latin America, and other regions combined.The original coverage targeted 14 African, 41 Asian, 42 European, and 4 Latin American languages.Source: Geographic-Based Contribution Assessment · p. 7, Figure 4 · p. 7

2.2 Annotation tasks

• Aya’s annotation workflow comprised original annotations, re-annotations, and annotation feedback following a find-fix-verify paradigm.These tasks respectively created new examples, improved existing examples, and assessed contribution quality.Source: §2.2 · p. 8

2.2.1 Original Annotations

• Original annotations elicited human-written prompt-completion pairs intended to represent contributors’ languages, cultures, literature, histories, and regions.The task addressed concerns that multilingual model generations can reflect Western cultural and representation biases [Yuan et al., 2021].Source: §2.2.1 · p. 8

2.2.2 Re-Annotations

• Re-annotations let contributors edit existing examples, familiarizing them with instruction formats while enabling human evaluation of datasets spanning question answering, summarization, paraphrasing, and translation.The platform initially populated re-annotation tasks with 1M translated prompt-completion pairs sampled from 19 public sources across 114 languages using NLLB 3.3B.Source: §2.2.2 · pp. 8–9

2.2.3 Annotation Feedback

• Peer reviews produced an Aya quality score, with original submissions reviewed through re-annotations and re-annotations reviewed through annotation feedback.This sequential pipeline was designed to evaluate and enhance collected data, while annotator engagement averaged 1.3 sessions and five contribution days.Source: §2.2.3 · pp. 10–11, Figure 6 · p. 10

2.3 Criteria for Inclusion in Aya Dataset

• The final Aya Dataset released original annotations and only re-annotations differing by at least five character-level Levenshtein edits, including languages with at least 50 contributions.These criteria excluded duplicate edits while balancing data quality, language coverage, and human-resource constraints.Source: §2.3 · p. 11

3.1 Statistics

• 204,114 instances span 65 languages: 22 high-resource, 12 mid-resource, and 31 low-resource.The dataset was collected through the Aya Annotation Platform.Source: §3.1 · p. 11

3.2 Length of Aya Dataset

• 25% longer completions followed re-annotation on average, with Aya original annotations increasing by 40%; combined prompt-completion length correlated with approval at 0.27.Re-annotation increased average completion length across all data sources, while longer prompts and completions were positively associated with perceived quality.Source: §3.2 · pp. 12–13, Figure 7 · p. 12, Figure 9 · p. 13

3.3 Annotator Skew

• Language contributions ranged from 14,597 Malagasy instances to 79 Kurdish instances, while 12 languages received all examples from their five most active annotators.The median number of annotators per language was 15, but Sindhi had one active annotator and English and Portuguese had over 80; Zulu and Sindhi each had one annotator responsible for all contributions.Source: §3.3 · pp. 14–15, Figure 10 · p. 15

3.4 The impact of introducing the Aya Score

• Introducing the Aya Score produced a marked uptick in completion length and increased the volume of original annotations and re-annotations.The score encouraged edits that transformed short answers into full sentences or paragraphs.Source: §3.4 · pp. 15–16, Figure 11 · p. 16

4 Aya Collection

• The Aya Collection provides researchers with a large multilingual corpus restricted to datasets permitting manipulation and redistribution, with licenses reported where possible.12Source: §4 · p. 16
• Its three sources are templated data, translated data, and the Aya Dataset, the collection’s only dataset human-annotated in its entirety.Templated data expand existing datasets through fluent-speaker-created templates, while the Aya Dataset is released as a subset of the overall collection.Source: §4 · p. 16
• Dataset selection prioritized diverse tasks, natural complete sentences, and prompt-completion suitability, excluding datasets likely to produce single-word answers or noisy outputs.Source: §4 · pp. 16–17

4.1 Templating Existing Datasets

• Aya uniquely uses fluent speakers to create language-specific prompts for expanding existing datasets, rather than relying only on English prompts or automatic translation.Using PromptSource, community members created templates in languages they knew; prompts could be monolingual or code-mixed, with English instructions also annotated for non-English datasets.Source: §4.1 · p. 17

4.2 Automatic Translation

• Nineteen high-quality datasets were translated with NLLB’s 3.3B-parameter model into 101 languages and 114 dialects, with translations prioritized for task diversity and completion length.The translated datasets are open-sourced in the Aya Collection and draw from xP3, the Flan Collection, Dolly, SODA, Mintaka, and additional sources.Source: §4 · p. 16, §4.2 · p. 18
• 48.17 mean ChrF++ on FLORES was reported for English-out translations, but Standard Arabic in Latin script and Minangkabau in Arabic script were excluded after significant errors were found.The score ranged from 10.9 to 69.6 across languages; translated Dolly also supplies a manually curated conversational instruction resource for languages lacking such datasets.Source: §4.2 · p. 18, Figure 13 · p. 18

5.1 Statistics

• 513M individual instances across 114 languages comprise the final Aya Collection, including 44 multilingual or non-English templated datasets and 19 translated datasets.The paper’s collection statistics exclude the Aya Dataset; including it raises the total language count to 115.Source: §5.1 · p. 19
• The collection spans Question Answering, Natural Language Generation, and Text Classification, subdivided into 11 finer-grained task types.NLG includes summarization, translation, paraphrasing, dialogue generation, and text simplification; classification includes six finer-grained types, while QA retains its main task label.Source: §5.1 · p. 19
• Including translated datasets reduces language disparities, while templated data maintains reasonable coverage for languages with fewer Wikipedia pages without directly comparing incompatible absolute values.The analysis compares patterns of data scarcity between collection instances and Wikipedia presence rather than their absolute values.Source: §5.1 · p. 20, Figure 14 · p. 20
• Low-resource languages occur at both ends of prompt and completion-length distributions, with Somali and Amharic showing longer average completions than some medium- or high-resource languages.The authors attribute this coverage to participants identifying and templating datasets in their own languages.Source: §5.1 · p. 20, Figure 15 · p. 21

5.2 Quality Assessment of All Different Data Sources

• 0.81 average approval ratio made Aya original annotations the highest-quality source group, compared with approximately 0.50 for xP3 among datasets receiving at least 20 votes.The quality analysis covered 40 datasets, and all translated data plus Original Annotations were above the 0.5 average threshold.Source: §5.2 · p. 21, Figure 16 · p. 22

• The suite fills a gap in multilingual evaluation by targeting open-ended generation tasks such as brainstorming, planning, and unstructured long-form responses.Source: §6 · p. 22
• Its design combines human-curated examples in limited languages, automatically translated examples across more languages, and human-post-edited translations in selected languages.The two primary sources are culturally curated Aya annotations and carefully selected, human-written Dolly prompts.Source: §6 · p. 22

aya-human-annotated test set

• The aya-human-annotated test set contains 1,750 instances, with 250 examples each across seven languages selected for resource level, script, and language-family diversity.The languages are English, Portuguese, Simplified Chinese, Standard Arabic, Telugu, Turkish, and Yoruba.Source: §6 · pp. 22–23

dolly-machine-translated test set

• The dolly-machine-translated test set retains 200 culturally screened Dolly prompts and translates them into 101 languages and dialects, covering 115 dialects with English included.Two reviewers screened 500 prompts, with a third resolving disagreements, to reduce dependence on culturally or geographically specific knowledge.Source: §6 · p. 23

dolly-human-edited test set

• The dolly-human-edited test set post-edits translated prompts in Arabic, Hindi, Spanish, French, Serbian, and Russian to improve interpretability when automatic translations become nonsensical or transfer poorly.The fish-name example shows that literal translation can erase the intended task even when the source prompt is clear.Source: §6 · p. 23

Limitations and recommended use

• The machine-translated set broadens coverage to 101 languages and 114 dialects but may misestimate ability when translation quality is inadequate, so it should be paired with human-edited or human-annotated sets.The Aya-human-annotated set covers seven languages and is entirely created by proficient target-language speakers.Source: §6 · pp. 23–24

Motivation

• Narrow research networks and undervalued data creators can constrain innovation and compromise linguistic diversity, motivating participatory collaboration with language speakers and local communities.The section emphasizes that data is connected to the people who produce, prepare, and distribute it, rather than being an exploitable natural resource.Source: §7 · p. 24

Participatory Research

• Aya’s cross-institutional global collaboration enabled linguistic data collection and community engagement considered crucial for effective language technologies [Joshi et al., 2019; ∀et al., 2020].The project is presented as an example of participatory research conducted over one year.Source: §7 · p. 24

Fluid Ownership and Growth

• Aya used decentralized, democratic collaboration with fluid leadership, flexible roles, asynchronous communication, and open recorded meetings to support participation across expertise and time zones.Members could become Language Ambassadors or mentors as their interests and experience evolved during the project.Source: §7 · p. 24

Inclusion and Access

• The open Aya interface reduced academic gatekeeping, involving non-English speakers and people without formal credentials whose language expertise expanded the representativeness of the data.Participants also gained exposure to machine learning, while expertise in a language was treated as distinct from machine-learning expertise.Source: §7 · pp. 24–25

Who Participated in Aya

• Contributors joined primarily through community, identity, and social-justice motivations, advocating for linguistic inclusion and preserving culturally specific language practices.Examples include planned Malagasy voice-to-text hackathons and efforts to include Telugu Sathakam poetry in Aya [Castaldo, 2023].Source: §7 · p. 25

Participation and Constraints

• Contributor relationships and debates over language representation were central to Aya’s success and data provenance, while conflict, outages, weak connectivity, and shipping barriers disrupted participation.These conditions shaped both contributors’ experiences and project progress, making the dataset’s breadth, provenance, shortcomings, and living history inseparable from its production context.Source: §7 · pp. 25–26

8.1 Multilingual datasets

• Multilingual datasets include unstructured pre-training corpora, parallel machine-translation collections, and smaller task-specific resources that trade scale for targeted quality.Large corpora often require extensive cleaning, whereas datasets such as XCOPA [Ponti et al., 2020] and XNLI [Conneau et al., 2018] target specific cross-lingual capabilities.Source: §8.1 · p. 28

8.2 Instruction-tuning datasets

• Instruction-tuning datasets combine human-curated, templatized, and synthetic instruction-response pairs, including broad meta-datasets and application-specific collections.Synthetic resources include Self-Instruct [Wang et al., 2022b], Alpaca, and Guanaco [Joseph Cheung, 2023], while manual efforts include Open Assistant [Köpf et al., 2023], Dolly, and LIMA [Zhou et al., 2023].Source: §8.2 · p. 29

8.3 Multilingual Instruction-Tuning Datasets

• Prior work extended instruction-finetuned datasets beyond English through translation, templates, or human curation, whereas Aya uses unique human-curated datasets and non-English instructions across 65 languages.xP3 uses the same templates across languages and exclusively English task instructions, while machine-translated prompts can lack variability and cultural nuance. Aya’s non-English instructions are created by human annotators.Source: §8.3 · pp. 29–30

8.4 Participatory Research in Machine Learning

• Participatory machine-learning initiatives have addressed regional languages, specific tasks, or image data, while Aya adopts a global, text-focused collaboration spanning diverse languages and continents.Earlier NLP efforts include Latin American Spanish dialects and 30 under-represented African languages [∀et al., 2020], whereas open-science programs such as BigScience produced BLOOM [Scao et al., 2022a] and ROOTS.Source: §8.4 · pp. 30–31

Language and dialect coverage

• The Aya Dataset and Collection cover 65 and 114 languages, yet 93% of the world’s languages remain unused for LLM training, with many dialects and code-switching patterns absent.Regional Malay dialects are missing, languages were isolated for classification, and programming languages were excluded from coverage.Source: §9 · pp. 31–32

Uneven distribution of contributions

• Annotator activity and language contributions are highly uneven, creating long-tailed participation, unequal dataset sizes, and limited annotator diversity in some languages.A few frequent contributors account for much of the data in certain languages, increasing the risk of narrow representation.Source: §9 · p. 32

Cultural or personal bias

• Skewed contributors and source availability can bias annotations toward particular cultural viewpoints and news-style language rather than everyday speech.For example, French data may emphasize France over Francophone communities elsewhere, while African-language data may overrepresent news text.Source: §9 · p. 32

Gendered pronouns and formality distinctions

• Gendered pronouns, honorific choices, and inconsistent standardization require context-sensitive responses and can encode assumptions about gender, status, identity, and regional language norms.Existing gendered annotations may remain unflagged, while annotators resolved formality choices according to speech-community norms; Portuguese variation illustrates regional differences in grammar, spelling, vocabulary, and formality.Source: §9 · pp. 32–33

Toxic or offensive speech

• Toxic or offensive content may remain because the platform lacked dedicated flags and not every entry was audited, despite human verification, peer review, and no reported incidents.Offensiveness can depend on annotator or political context, so mitigation does not guarantee complete removal.Source: §9 · p. 33

Accounting for mislabeled data

• The platform could not relabel languages, and some dataset entries may be miscategorized or fail to use instruction-style formats.These issues may arise despite annotators’ adherence to directions and extensive manual auditing.Source: §9 · p. 33

10 Conclusion

• Aya’s participatory research involved people from different countries, ages, and levels of NLP familiarity, illustrating opportunities for collaboration with social scientists.The authors connect this model to emerging open-science norms.Source: §10 · p. 34
• Aya Dataset contains 204,114 human-curated prompt-completion pairs across 65 languages, created by 2,997 collaborators from 119 countries over eight months.The dataset is released as an open-source multilingual instruction-style resource.Source: §10 · p. 34
• Aya Collection comprises 513M prompt-completion pairs across 114 languages, including translations into 101 languages, making it the largest multilingual instruction-finetuning collection today.It contains 44 instruction-style datasets transformed from existing NLP datasets and expands coverage for low-resource languages.Source: §10 · p. 34
• Aya also releases a human-curated Evaluation Suite covering 13 languages and translated prompts covering 101 languages.The suite complements the multilingual training resources with evaluation examples and prompts.Source: §10 · p. 34
• The open-sourced Aya Annotation Platform is intended to help communities continue multilingual data collection and connect speakers of low-resource languages worldwide.The platform is part of a broader effort to change how multilingual NLP data is created.Source: §10 · p. 34

A Aya Annotation Platform

• The annotation platform and gamification strategy were designed together to support high-quality curation of the Aya Dataset.The platform collects demographic information and annotator language preferences during onboarding, which determine later language options in the interface.Source: A Aya Annotation Platform · p. 68, Figure 17 · p. 68

A.1 Engagement Strategies

• Gamification combined points, social-media recognition, leaderboards, mini-challenges, real-time feedback, and avatars to support engagement, data quality, and community.These mechanisms aimed to create friendly competition, collective achievement goals, positive reinforcement, and a more enjoyable crowdsourcing experience.Source: §A.1 · p. 68
• Contributors reaching 500, 1,000, and 5,000 contributions received milestone-specific certificates, Aya apparel, and documentation recognition.The tiered rewards recognized sustained contribution and incentivized continued participation.Source: §A.1 · p. 68
• The Aya Discord Bot recognized the daily top 10 contributors, aggregated regional totals, and reported the remaining collection time.These updates complemented leaderboard-based recognition with recurring project-status information.Source: §A.1 · p. 68
• Figure 18’s regular progression snapshots helped annotators see the dataset grow across all languages.The Discord messages presented daily statistics and top-annotator information to the contributor community.Source: §A.1 · p. 69, Figure 18 · p. 69
• The Aya leaderboard ranked annotators through daily, weekly, language-specific, and overall views based on their Aya score.These views made contribution standings visible at multiple timescales and aggregation levels.Source: Figure 19 · p. 69

A.2 Accessibility of Registration Tools

• Registration-tool accessibility and popularity differed by country, influencing where Aya UI users joined from.Figure 20 compares Discord- and Gmail-based registration percentages across the top 10 countries.Source: §A.2 · p. 70, Figure 20 · p. 70
• After Google SSO was introduced, the number of registered users increased significantly in several new countries.Figure 21 reports the share of Aya UI users who registered after Google SSO was added.Source: §A.2 · p. 70, Figure 21 · p. 70

B.1 Reviewing Annotators

• Reviewers compared original and re-annotated prompts and completions on a 1–5 scale, with scores below 5 allowing edits and additional feedback.The scale ranged from “Much worse” to “Much better” than the original annotation.Source: §B.1 · pp. 70–71

B.2 Aya Score

• The Aya Score ranked annotators using re-annotation activity, submission counts, peer-reviewed quality, and positive feedback to encourage high-quality contributions.Its stated rationale was to foster competition and improve the overall quality of the Aya Dataset.Source: §B.2 · p. 71

B.3 Annotation Guidelines

• Annotation guidelines required error-free, clear, correct, coherent, reasonably sized text, with original annotations not generated by other language models.Annotators used thumbs-up or thumbs-down before editing and were encouraged to expand short completions with details and context.Source: §B.3 · pp. 71–72, Re-Annotations · p. 72

C.1 Division by Regions

• Languages were administered across Africa, Asia, Europe, and Latin America, comprising 14, 41, 42, and 4 languages respectively, with exceptions and cross-region assignments for some languages.Regional Leads coordinated Language Ambassadors and recruited fluent speakers within each area.Source: §C.1 · p. 72

C.2 Language Ambassadors

• Eighty-four native-fluent Language Ambassadors broadened participation and addressed language-specific, cultural, regional, and data-quality issues.They also helped select suitable data sources and received asynchronous and weekly online support.Source: §C.2 · p. 72

C.3 Regional Leads

• Six voluntary Regional Leads, selected from their regions, chose Language Ambassadors, recruited annotators, and helped sustain regional participation.There were two Regional Leads for Latin America, one for Africa, one for Asia, and two for Europe.Source: §C.3 · p. 73

D.1 Platforms

• A Discord server supported internal coordination, while X, LinkedIn, WhatsApp, and Facebook enabled culturally and regionally flexible external outreach.Channels covered project functions, regions, languages, and Language Ambassadors.Source: §D.1 · p. 73

D.2 Meetings

• Recurring meetings combined weekly, bi-weekly, and monthly updates, onboarding, troubleshooting, technical discussion, language-specific co-working, datathons, and synchronous annotation support.Meetings included Regional Leads and Language Ambassadors, new contributors, Regional Leads, technical updates, and language-specific sessions.Source: §D.2 · pp. 73–74

E Language Groupings

• Languages are categorized as lower-, mid-, or higher-resourced using NLP-resource clusters from Joshi et al. [2020], serving as an imperfect aggregation proxy for data availability.The authors caution that language varieties cannot be universally classified along this single dimension.Source: E Language Groupings · p. 74

F.1 Annotators

• Annotators were recruited for native or professional proficiency in the required languages and came from varied social, educational, and employment backgrounds.Socio-demographic information was gathered informally through self-reporting rather than formal statistical collection.Source: §F.1 · p. 75
• Annotators were paid 30 CAD per hour, the standard rate for Cohere annotators performing highly complex tasks.Source: §F.1 · p. 75

F.2 Annotation Process

• The annotation process used flexible schedules, Google Sheets, and 3–4 annotators per task, with a second annotator verifying each post-edit and proposing final revisions when needed.Annotators received an introduction, could raise questions in Slack, and were encouraged to report recurring errors and difficult decisions.Source: §F.2 · p. 75

F.3 Instructions

• For machine-translated prompts, annotators were instructed to preserve the English instruction exactly, ensure grammatical correctness, and remove overly literal “Translationese.”They first judged whether each translation was acceptable, then minimally edited unacceptable prompts and confirmed that the three criteria were satisfied.Source: §F.3 · p. 76
• Acceptable prompts were left unchanged and confirmed against the criteria, while prompts requiring revision were edited minimally in a dedicated field.Source: §F.3 · p. 76

F Post-Editing the dolly-machine-translated Test Set

• Each prompt was edited by one annotator and verified by another, with discussion and further editing as necessary; three to four editors collaborated on each language.Source: §F.3 · p. 76

F.4 Post-Editing Effort

• At least 41% of prompts required editing in every language, indicating that translation errors were common in the dolly-machine-translated test set.Post-editing effort is reported using edited-prompt percentage, HTER, and HChrF; higher HTER and lower HChrF indicate greater effort.Source: §F.4 · p. 76, Table 7 · p. 77
• Russian required the most post-editing, averaging 37.43 HTER, while French had similar edited-prompt coverage but only 5.56 HTER.Edited prompts comprised 86.5% for Russian and 84.5% for French.Source: §F.4 · p. 77, Table 7 · p. 77

G Translation using NLLB

• NLLB translations of full paragraphs could terminate abruptly because the model was trained on single-sentence pairs, so paragraphs were sentence-tokenized and reassembled after translation.Samples containing <unk> tokens introduced by the NLLB tokenizer or model were filtered to avoid translation artefacts.Source: G Translation using NLLB · p. 77

G.1 Translation Quality of NLLB

• Figure 22 evaluates NLLB translation quality with ChrF++ on FLORES for Aya target languages covered by NLLB, grouped by resourcefulness according to.The scores were extracted for the dense 3.3B model from the fairseq NLLB README.Source: §G.1 · p. 77, Figure 22 · p. 78

H Additional Figures

• Additional figures report dataset-level approval ratios and prompt/completion length distributions before and after re-annotation, across language categories and individual languages.These summaries are provided in Figures 23–26.Source: Figure 23 · p. 79, Figure 24 · p. 79, Figure 25 · p. 80

I Additional Tables

• Additional tables catalogue Aya Collection datasets by templated or translated origin, language coverage, lengths, licenses, tasks, and task-taxonomy categories.Tables 8–12 list templated datasets, translated datasets, their task taxonomies, and xP3 datasets; the taxonomy includes broad tasks such as question answering, natural language generation, and text classification with finer-grained categories where available.Source: Table 8 · p. 83, Table 9 · p. 84, Table 10 · p. 85

J Data Cards

• The Aya Dataset contains 204,114 annotated prompt-completion pairs in 65 languages, curated by 2,007 contributors from 110 countries through volunteer annotations and quality assessment.The annotations were collected on the Aya Annotation Platform from June through December 2023.Source: J Data Cards · p. 88
• The data cards record the three resources’ sources, platform, collection dates, versions, and maintenance status, with the Aya Dataset, Collection, and Evaluation Suite listed as actively maintained.The Dataset and Collection use original or existing NLP datasets collected through the Aya Annotation Platform from June to December 2023; the Collection and Evaluation Suite state that no updates are planned.Source: J Data Cards · pp. 88–89
• The Aya Collection combines templates from 44 datasets with translations of 19 instruction datasets into 101 languages, yielding 513,579,625 prompt-completion instances across 114 languages.Its curation uses crowd-sourced templating and automatic translation, with languages aligned to mT5, annotator language lists, and the NLLB model.Source: J Data Cards · p. 89
• The Aya Evaluation Suite contains 25,750 open-ended conversation prompts covering 101 languages across human-annotated, machine-translated, and human-edited subsets.The subsets include 250 original prompts in 7 languages, 200 Dolly prompts translated into 101 languages, and 200 translated prompts post-edited for 6 languages.Source: J Data Cards · p. 90

K Aya Collection Templates

• The cited inventory includes AfriSenti-inst, Amharic QA [Abedissa et al., 2023], Hindi-article-summarization [ganeshjcs, 2023a], IndicSentiment-inst [el2e10, 2023a], IndicXParaphrase-inst [el2e10, 2023b; SuryaKrishna02, 2023d], and MasakhaNEWS-inst.Source: K Aya Collection Templates · pp. 91, 97, 100, 104
• The collection includes translation templates, including prompts asking for Ligurian translations of supplied text.A related template explicitly asks, “Can you translate this text to Ligurian?” and specifies the Ligurian translation as the response.Source: K Aya Collection Templates · p. 113
• Story-generation templates ask models to generate titles from stories or stories from supplied characters, words, and titles, including Tamil-language prompts.The templates include both title generation and short-story generation formats.Source: K Aya Collection Templates · p. 114
• Additional templates support Telugu riddles and Ukrainian grammatical-error correction, with positive corrected-text and negative no-error targets.The Ukrainian examples use target_pos for corrected text and target_neg for sentences containing no errors.Source: K Aya Collection Templates · pp. 114, 118
• Aya Collection templates cover multilingual named-entity recognition using PER, ORG, and LOC categories while excluding nationalities and requiring JSON outputs with entity spans.The templates include fields such as TypeName, Text, Start, and End, and instruct models to follow example structures.Source: K Aya Collection Templates · p. 120
• Aya Collection templates instruct models to identify named entities using only PER, ORG, and LOC categories, while excluding nationalities from locations and organizations.The instructions explain that organizations may represent other groups of people.Source: K Aya Collection Templates · pp. 121–122
• The templates require outputs in JSON format with fields for the entity type, text, and start and end indices.The specified structure uses the placeholders "TypeName", "Text", "Start", and "End".Source: K Aya Collection Templates · pp. 121–122
• The applied template list includes Urdu-News-Gen-Article [AhmadMustafa, 2023a], Urdu-News-Category-Class [AhmadMustafa, 2023b], Urdu-News-Gen-Headline [AhmadMustafa, 2023c], Wiki-split-inst [Botha et al., 2018], and Xlel_wd-inst.Table 13 presents the Aya Collection’s applied templates for each templated dataset.Source: K Aya Collection Templates · pp. 122–123, Table 13 · p. 124