Ochieng Millicent

Research Scientist @ Microsoft | Multilingual NLP . Generative AI · Equitable AI | Masakhane & WiDS Ambassador

[Ochieng Millicent @iam_OchiengM]

Excited to see our work featured in the latest Microsoft Research Focus edition alongside other great research.
Our paper, Reasoning Beyond Labels, explores how LLMs interpret sentiment in multilingual, culturally nuanced communication and why evaluation must go beyond accuracy!

[Microsoft Research @MSFTResearch]

In the latest Research Focus: LLM sentiment in cultural context, robot assembly learning, smarter AI agents, verified Rust code, and CHI 2026. https://t.co/JDW828G95X

[AfricaNLP @Africa_NLP]

A big thank you to our sponsors

[AfricaNLP @Africa_NLP]

Thank you to the organizers for making this edition possible

[AfricaNLP @Africa_NLP]

Ongoing panel discussion @Africa_NLP

Theme: What will it take for multilingual multimodal LLMs to work for Africa?

Panelists: @AtnafuLambebo @francoisrmeyer @IfeAdebara John Quinn
Moderator: @iam_OchiengM

[Ochieng Millicent @iam_OchiengM]

@davlanade @SIC_Saar Congratulations, David 🎊

[Ochieng Millicent @iam_OchiengM]

The award is here, and I had to share. 🏆
Our paper IrokoBench won the Outstanding Paper Award at #NAACL2025!
Proud to be part of this work advancing multilingual LLM evaluation across 17 African languages.
Huge thanks to @davlanade for visionary leadership.
#AfricanLanguages

[Ochieng Millicent @iam_OchiengM]

🎉 Just had an incredible experience attending Conference on Neural Information Processing Systems (NeurIPS 2024)! 🎉 #NeurIPSConference - via #Whova event app

[Ochieng Millicent @iam_OchiengM]

@alienelf Congratulations 🎊 👏

[Ochieng Millicent @iam_OchiengM]

@UR_aceds @CharlesRuranga @jbatamuliza @Uni_Rwanda @Ace2Africa @rbarwanda Amazing 👏

[Cohere Labs @Cohere_Labs]

Led by: @davlanade, @JessicaOjo__ , @Israel_Abebe , Jian Yun Zhuang, @alabi_jesujoba, @zodiacJRH, @iam_OchiengM, @sarahookr, @AndieBukula, En-Shiun Annie Lee, Chiamaka Chukwuneke, @happybuzaaba1, @banQomania, @jonathanmukiibi, @SalomonKabongo, @yuehgoh, Lolwethu Ndolela...

[Lydia Nishimwe @LydiaNishimwe]

@iam_OchiengM You can read the paper here https://t.co/edGspNGpyM

[Lydia Nishimwe @LydiaNishimwe]

Special thanks to @iam_OchiengM and the Microsoft AI for Good team for inviting me!

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

Benchmark is now on @huggingface collection
https://t.co/Iwy0cQFCLy

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

A big thank you to @OpenAI @CohereForAI and @Oracle for the compute/API credits. Also, we thank @haileysch__ and @lintangsutawika for their assistance with the EleutherAI lm-eval tool.

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

This is a great collaboration with many authors @JessicaOjo__ @Israel_Abebe , @alabi_jesujoba @zodiacJRH @iam_OchiengM @sarahookr @AndieBukula @happybuzaaba1 @banQomania @jonathanmukiibi @SalomonKabongo @yuehgoh @RoowyM @Shmuhammadd @NanaYaaSally @SambSokhar

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

In few-shot evaluation, LLaMa 3 70B significantly benefits from few-shot examples for AfriMMLU and AfriXNLI but did not for AfriMGSM since it's only able to reason effectively on maths in English.

GPT-4o consistently improves in performance with additional few-shot examples.

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

GPT-4o is the best across all tasks for native African languages, however, the performance is worse for eng & fra, where GPT-4-Turbo is more than +9.0 better. Aya-101 is the best open-model, but in the translate-test setting LLaMa 3 70B is better since it's more English-centric

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

We provide zero/few-shot evaluation of the performance of 14 LLMs (open weight models and closed models) in two settings leveraging lm-eval:
1) in-language and
2) translate-test (where test sets were automatically translated to English using NLLB-200-3B).

[David Ifeoluwa Adelani 🇳🇬 @davlanade]

We cover 18 languages in our benchmark, 16 native African languages (amh, ewe, hau, ibo, kin, lin, lug, orm, sna, sot, swa, twi, wol, xho, yor, & zul) , and two European languages (eng and fra) translated from MGSM, MMLU and XNLI datasets.