pith. sign in

arxiv: 2405.15032 · v2 · pith:DMRFVOC2new · submitted 2024-05-23 · 💻 cs.CL

Aya 23: Open Weight Releases to Further Multilingual Progress

classification 💻 cs.CL
keywords multilinguallanguagesmodelmodelslanguageexpandinglikeopen
0
0 comments X
read the original abstract

This technical report introduces Aya 23, a family of multilingual language models. Aya 23 builds on the recent release of the Aya model (\"Ust\"un et al., 2024), focusing on pairing a highly performant pre-trained model with the recently released Aya collection (Singh et al., 2024). The result is a powerful multilingual large language model serving 23 languages, expanding state-of-art language modeling capabilities to approximately half of the world's population. The Aya model covered 101 languages whereas Aya 23 is an experiment in depth vs breadth, exploring the impact of allocating more capacity to fewer languages that are included during pre-training. 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 on an extensive range of discriminative and generative tasks. We release the open weights for both the 8B and 35B models as part of our continued commitment for expanding access to multilingual progress.

This paper has not been read by Pith yet.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. When Similar Means Different: Evaluating LLMs on Arabic--Hebrew Cognates

    cs.CL 2026-06 unverdicted novelty 7.0

    LLMs achieve high accuracy on true Arabic-Hebrew cognates but drop sharply on false friends and loanwords due to surface-form reliance, with only modest gains from sentence context.

  2. SomaliBench Eval: Measuring English-to-Somali Refusal Gaps in Open-Weight Language Models

    cs.CL 2026-05 unverdicted novelty 7.0

    SomaliBench finds large English-to-Somali refusal gaps (0.38 to 0.90) across Llama-3.1-8B, Gemma-2-9B, Qwen-2.5-7B, and Aya-23-8B, with many Somali responses being unclear rather than compliant.

  3. Vocab Diet: Reshaping the Vocabulary of LLMs via Vector Arithmetic

    cs.CL 2025-10 conditional novelty 7.0

    LLMs can compose surface-form tokens from base embeddings plus learned transformation vectors, freeing 10-40% of vocabulary slots while expanding coverage and preserving downstream performance across five languages.

  4. Sentence-Level Contextual Entrainment in Large Language Models

    cs.CL 2026-06 unverdicted novelty 6.0

    Sentence-level contextual entrainment exists across LLMs, weakens with scale, and is localized to 2-4% of attention heads whose deactivation removes the effect without performance loss.

  5. Speaking in Self-Assessing Tongues: On the Verbalized Confidence of LLMs in Machine Translation

    cs.CL 2026-06 unverdicted novelty 5.0

    Empirical study finds verbalized per-token confidence methods in LLMs for MT perform similarly to internal signals on error detection and calibration but show little correlation.

  6. Patching LLM Like Software: A Lightweight Method for Improving Safety Policy in Large Language Models

    cs.AI 2025-11 unverdicted novelty 5.0

    Prepending a compact learnable prefix to LLMs produces safety gains comparable to next-generation aligned models while preserving fluency and adding negligible parameters.

  7. Optimizing Korean-Centric LLMs via Token Pruning

    cs.CL 2026-04 unverdicted novelty 4.0

    Token pruning of non-Korean vocabulary in LLMs improves generation stability and often boosts machine translation on Korean tasks while cutting vocabulary size substantially.

  8. Multilingual Vision-Language Models, A Survey

    cs.CL 2025-09 accept novelty 3.0

    The survey identifies a key tension in multilingual vision-language models between language neutrality via contrastive learning and cultural awareness via diverse data, with most benchmarks relying on translation-base...