We are committed to making meaningful progress in machine learning research through open collaboration. Follow this 🧵to stay on top of our research contributions.
Today, we’re launching Aya, a new open-source, massively multilingual LLM & dataset to help support under-represented languages. Aya outperforms existing open-source models and covers 101 different languages – more than double covered by previous models.
Today, we launch Aya 23, a state-of-art multilingual 8B and 35B open weights release.
Aya 23 pairs a highly performant pre-trained model with the recent Aya dataset, making multilingual generative AI breakthroughs accessible to the research community. 🌍
An exciting day for ! We believe in changing where, how, and by whom research is done. Today we launch our Scholars Program — an opportunity to work with some of the best researchers in the world. Your journey starts here.
We are excited to launch C4AI Command-R, a 35 billion parameter model weights release designed to make generative AI breakthroughs accessible to the research community. 🎉
Congratulations to head of Cohere For AI and VP of Research at
@cohere
,
@sarahookr
, for being recognized as part of
@TIME
's list TIME100 AI of 2024!
#time100ai
Today we officially launched ! We are very excited to be led by the brilliant
@sarahookr
as we aim to focus on collaborating on open source research and creating more points of entry into machine learning research. Want to join in?
We are excited to launch the next cohort of the Cohere For AI Scholars Program — an opportunity to work with some of the best researchers in the world. 🎉 The Scholars Program is designed to change where, how, and by whom research is done. Apply by Aug 30!
Announcing C4AI Command R+ open weights, a state-of-the-art 104B LLM with RAG, tooling and multilingual in 10 languages.
This release builds on our 35B and is a part of our commitment to make AI breakthroughs accessible to the research community. 🎉
What is “good data?”👩🔬
Our recent paper tackles this question via data pruning! We explore several metrics for measuring LLM pretraining data and finds that we can remove up to 70% of pretraining data while achieving better test set performance.
📜
Today we are thrilled to launch the second cohort of our Cohere For AI Scholars Program — an opportunity to work with some of the best researchers in the world. This Program is designed to change where, how, and by whom research is done. Apply by Sept. 11!
Less than 24 hours after release, C4AI Command-R claims the
#1
spot on the Hugging Face leaderboard!
We launched with the goal of making generative AI breakthroughs accessible to the research community - so exciting to see such a positive response. 🔥
We’re celebrating our first year of Cohere For AI and the moments that made this year special! Here’s a look back at the highlights at Cohere For AI on our first anniversary! 🧵
We’re excited to launch the Cohere For AI Research Grant Program 🎉
C4AI research grants are designed to support academic partners conduct research with the goal of releasing a peer-reviewed scientific artifact or involving a data-for-good project 🔬
We’re excited to launch Roads to Research 🛣️- a new program to showcase all it takes to be an ML researcher, with focus on elements of the process that are less often spotlighted.
First up:
@sarahookr
on “Your journey into research: lessons to live by”
We have released a 4-bit quantized version of C4AI Command R+, a 104B model weights release designed to make generative AI breakthroughs accessible to the research community. ✨
C4AI Command R is now available on
@huggingface
Spaces to try out. ✨
35B open weights with multilingual evaluation in 10 languages, reasoning, summarization, and question answering. 🔥
We have released a 4 bit quantized version of C4AI Command-R model, a 35 billion parameter model weights release designed to make generative AI breakthroughs accessible to the research community. 🎉
📣Announcing our new cross-institutional collaboration.
We've brought together researchers invested in improving multilingual benchmarks. We're starting with MMLU, a heavily translated dataset used for multilingual evals that doesn't capture cultural nuances.
Let's address this
Our community-led Beginners in Research-Driven Studies (BIRDS) group is kicking off it’s first mini-cohort learning group focused on CUDA Programming for Beginners, beginning on Friday, April 5th 🎉
Many communities have been left unsupported due to the language limitations of previous models. That’s why we’re open-sourcing both the Aya model and dataset, which includes more than 50 previously unserved languages, to ensure AI can serve a broad global audience. 🌍
Since launching Aya 1 week ago, our model has claimed the number 2 spot in the
@huggingface
leaderboard, and our datasets hold spots 1 & 2 for data. 🌏
Aya was always meant to be used, researched, & built upon by people worldwide - we're excited to see what you're working on! 🛠️
Aya Model paper wins a Best Paper Award at ACL, 2024.
We're honoured that our initiative to push the boundaries of multilingual AI and scientific practices has been so prestigiously recognized.
Thank you to everyone who has believed in Aya. 🌍
We have a community!
We are starting small, but we want to hear from you. You can find more details about our community at .
We will be holding the first meet and greet at the end of this month, so make sure to reach out by then.
As part of our commitment to accessibility, we are particularly proud of the 8B version of Aya 23, which is significantly smaller than previous releases making it more accessible for the research ecosystem. 🐁🐘
Try it here:
C4AI and our
@CohereAI
colleagues are getting ready to showcase our research and connect with our community at
#NeurIPS2022
! Here’s where you can find us at a glance. We’ll be at booth
#615
too, so please stop by and say hello! 👋
We are excited to announce the launch of the Data Provenance Initiative. ⛵ 🌊 🧭
A cross-institutional effort involving experts across 13 institutions to shine a spotlight on data transparency and attribution in AI. 🔍
🌐
We’re celebrating our second year of Cohere For AI with 🔟 ways in which we have delighted in exploring the unknown. 🎉Here’s a look back on C4AI highlights on our second-year anniversary! ✨
How did 3,000 independent researchers, across 119 countries come together to build a new state-of-art open-source multilingual model & dataset? 🌍🌿
The Journey of Aya, a 20-minute documentary, captures the breadth and humanity behind this work.
Dive deeper into the work with our 2 research papers. Our dataset research paper looks at the goal of making data sourcing for multilingual AI research more inclusive, allowing people from all around the world to have a hand in shaping and using LLMs. 📝
We’re incredibly excited to see 5 of our papers accepted into ACL today! 🎉We are so proud of all of the authors and collaborators who made these papers possible.
If you have been considering applying for the Scholars Program, now is the time to pull together your application materials and click submit!
Applications are due tomorrow, Monday, September 11th by the end of day, anywhere on earth! Apply now:
Curious about how language models manage to keep learning new things over time? We investigated continual pretraining in large language models on a diverse set of domains - our findings reveal new insights about knowledge transfer and forgetting. 🤔
📜
We invited C4AI community members to meet up at
@CohereAI
's London office for festive cheer and mulled wine to celebrate our first six months!🎄🍷. Happy holidays everyone!❄️
Elo has proven effective for dynamic games like chess and has recently seen widespread use for evaluating LLMs. But how reliable is it for evaluating static-skill entities like LLMs? 🤔♟️
We find scenarios where it’s not!
📜 Paper link:
Aya is now available on
@kaggle
models! 🎉
So cool to know that our work will now be available to the 17M learners, developers, and researchers in the Kaggle community. 🤯
One of the most exciting aspects of the full-time paid Scholars Program is that we will accept remote candidates from around the world.
We look forward to collaborating with early-career ML researchers across LatAm, Europe, and Africa! 4 days to apply.
The first meeting of our community's Interactive Reading Group is at the end of the week! Led by
@karinanguyen_
@AThatipelli
@NeelBhandari9
we'll be diving into Memorizing Transformers. Join our community + the conversation at
It has been an incredible year at C4AI – we are honored and humbled by all your support that has made this year special! Here’s a look back at our 2023 highlights! 🧵
Link to blog post:
Let’s make RLHF training more accessible and easier to implement. 🔥
Introducing RLOO in TRL - a GPU memory and wall clock time efficient RLHF training algorithm.
📝Learn more in this blog post from
@aahmadian_
and
@vwxyzjn
.
🌟 Calling all bright minds from South America 🌎✨
With just 3% of applications coming from South America, we’re looking for more Scholars Program applicants in South America! If you’re in search of an opportunity to develop your research skills, your journey starts here.
We have heard from some of you that it is hard to filter the huge Aya Collection if you only want a single language. 🌍
Given 101 languages and 513 million data points, this makes sense. :) Excited to share we now have a version split by language. 🥳
Applications are starting to roll in for our Scholars Program. 🎉
We would love to see more applicants, particularly from LatAm and Africa. 🌍
This is a remote-first, paid opportunity, with the goal of being one of the first doors to open for early-career ML researchers.
We are excited to launch the next cohort of the Cohere For AI Scholars Program — an opportunity to work with some of the best researchers in the world. 🎉 The Scholars Program is designed to change where, how, and by whom research is done. Apply by Aug 30!
It's time to level up contributions to the Aya project. 🚀Many languages have received fewer than 100 contributions.
Join us to make sure these languages are not left behind in the development of language AI!
Aya 23 is now available for our community. It can be used to experiment, explore + build on for research in 23 languages spoken by close to half of the world's population.
We’re excited to see what you do with it.✨
Try it out on
@huggingface
Spaces:
We're delighted to announce our new event series: Cohere For AI Fireside Chats.
Episode One features Samy Bengio, Senior Director of AI and Machine Learning Research at Apple, in conversation with
@sarahookr
.
👉 Registration is open now:
We're celebrating the new year by announcing our first speaker of 2023:
@colinraffel
🎉
Join us for a lively chat about Colin's research journey AND check off your new year's resolution to learn more from inspiring people. ✨
Register here!
We have completed 18% of the translation audits of our MMLU data subset into 39 languages. 🎉Thanks to everyone who has helped with this first push!
12 days left to join us and make MMLU work for the world 🌍
We open source the largest multilingual instruction fine tuned dataset to date, with 513 million prompts and completions covering 114 languages. This offers a large-scale repository of high-quality language data for developers and researchers.
How do we do more 🐘 with less 🐁?
In an era of ever larger models, work on efficiency is ever more important. This recent cross-institutional collaboration provides a survey of the field for practitioners and researchers alike ⚙️.
📜Learn more:
We’re excited to showcase our latest research and connect with our community during
@iclr_conf
! Here’s where you can find us and our
@CohereAI
colleagues throughout the conference.
Stop by the Cohere booth to say hello, chat with our team, and pick up some swag!
Our new Beginner ML Researchers group is gearing up for its kick-off event, Feb 23! Learn how to get involved +
@sarahookr
will share some tips & tricks.💡
Thanks to
@akankshanc
,
@krypticmouse
&
@rzsgrt
for organizing.
Join our community to join the fun!
The next event in our community mentorship series will be on Dec 8: Career c̶h̶o̶i̶c̶e̶ creation for non-standard candidates, with
@savvyRL
.
A big thank you to C4AI community organizers
@oohaijen
and
@jonas_kg
for hosting this event!
Register here:
Scholars Program applications are coming in from around the world, but we want to fill in more of our map!
Do you see your country represented yet?
Apply to the Cohere For AI Scholars program today:
Hello world!👋🌏Our community is made up of ML researchers from 74 countries who are in conversation with one another, supporting each other's initiatives, and learning from another's experiences.
We'd love to have you here as well! Learn more and join at
“Active inheritance is where you sample different parts of the problem you want to solve from diverse models. This diversity spurs interesting patterns, increasing the realm of possibility and quality that transcends any single model” -C4AI head,
@sarahookr
, for
@MLStreetTalk
A passion for mentorship, the challenge of working in academia in the era of scale, & guitar pedals:
@colinraffel
treated us to an insightful overview of his research journey during our first fireside chat of the year.
Catch up on the conversation:
It was a complete honour to welcome
@andreas_madsen
to speak on independent research and interpretability earlier this week! 🤩
For those who couldn't make it, we're excited to share a recording of this C4AI community-led talk.
Spirits are high at
@Khipu_AI
! ☀️
There's still time to say hello at the
@CohereAI
booth (503) and chat with our brilliant team about their research and all things NLP!
We're excited to host
@colinraffel
for our first speaker of the new year! 💙
Join us on January 12th to hear Colin share what he's learned along his research journey!
Register here:
One week until we welcome Samy Bengio, Senior Director of AI and Machine Learning Research at Apple, to be in conversation with
@sarahookr
.
Join us for our first Cohere For AI Fireside Chat by registering at
Catch the replay of
@Tim_Dettmers
fantastic "8-bit Methods for Efficient Deep Learning" Tech Talk!
Thanks to Tim for sharing his work and our keen audience for asking such great questions!
The
@aclmeeting2024
conference in Bangkok, Thailand is less than a week away and we are excited to share what our Cohere and C4AI teams will be presenting.🙌
Check out where we are throughout the conference here:
We're excited to host our first C4AI Technical Talk!
🔎Join us on September 21st for "Next Generation of Semantic Search" with
@Nils_Reimers
.
Sign up here 👉
Our community-led ML Theory Group is looking forward to hosting
@PetarV_93
, Research Scientist at
@GoogleDeepMind
, next week on Thursday, April 25th for a presentation on "Categorical Deep Learning. An Algebraic Theory of Architectures."
Learn more:
How does quantization affect multilingual LLMs? 🌍
For wide adoption, multilingual LLMs must be highly-performant *and* lightweight. 📈 🪶We analyze SOTA multilingual LLMs in 23 languages under various quantization techniques to find out.
📜
The Aya Model is a massively multilingual model that follows instructions in 101 languages. We significantly expand the size of available training data to address the linguistic inequality of recent NLP development & achieve state-of-the-art performance.
Is RLHF effective for aligning multilingual LLMs? 🤔
Our work studies multilingual preference optimization to train a new SOTA multilingual LLM, advancing the frontier of alignment techniques to 23 languages covering half the world’s population 🌎!
ICYMI 👀 The application process for our Scholars Program includes a takehome technical exercise and a personal statement. Become familiar with both if you plan to apply:
Questions? Join our info session on Oct 18:
On Wednesday we had the pleasure of gathering in Toronto to celebrate the 1 year anniversary of Cohere For AI.
Together we celebrated with food, drink, a photo booth, a champagne toast, and incredible conversations with the 100+ community members who were in attendance. 🍾✨🎉
Congrats to Tamil: the first language to surpass 10,000 contributions 💚
Sinhala also bumped up a category this week. 🥳
Still many Asian languages with less than 100 contributions. Let's turn that around!
Get started in your language
Safety isn't one-size-fits-all. It varies by culture, location and language, yet traditional alignment work often treats it as static.
Excited to introduce our new work on alignment that captures both local 🧧🎃🗿 and global 🌎 preferences!
📜
“Aya will serve as a wake-up call for industry and governments to consider language representation.”
@Joe_Castaldo
captures our year-long Aya initiative for the
@globeandmail
.
“Are you free right now?”
“I have to read this paper on LLM implicature performance”.
Humans effortlessly translate this response to "no". LLMs evidence clear limitations on implicature performance.
Learn more about our recent collaboration:
Why do current state-of-art language models only cover a handful of the world's 7000 languages, and what can we do about it? 🌎🌍🌏⁉️
Our latest primer explores this “language gap” in AI and offers policy & governance considerations to address it 📄✅
We want to thank all our collaborators on the Aya Project. This work brought together 3,000+ independent collaborators from 119 countries, making it one of the largest open science projects to date in the field of machine learning. 🎉
We are celebrating C4AI Research Scientist
@mziizm
, who has been nominated for Women in AI Netherlands Top 5 AI Researchers. 🎉
Congratulations, Marzieh, on this well-earned honour!
🎉 Milestone alert: the C4AI Community has hit the 1000+ members mark! 🎉
These 1000+ ML researchers hail from 95 countries and engage in discussions, attend meetings & events, and collaborate on research.
We're honoured to learn, grow, and explore with you all. Here's to YOU!
We are excited to announce the launch of the Data Provenance Initiative. ⛵ 🌊 🧭
A cross-institutional effort involving experts across 13 institutions to shine a spotlight on data transparency and attribution in AI.🔍
🌐
See behind the scenes in The Story of Aya, a 20-min. documentary featuring many of our collaborators that highlights the importance of progress in this field, and how this major research effort came together over the past year.
The Cohere For AI Scholars Program is an opportunity for a new class of machine learning talent to work alongside some of the best ML research & engineering expertise in the world.
Start your research journey. Apply to the Scholars Program today:
We're pleased to share that Aya 23 - 8B is now hosted and available via the Cohere API. 🔥
Our hope is that this will continue to make multilingual generative AI breakthroughs more accessible to the research community. 🌍
📣 Calling all Arabic speakers! On July 11 Aya Language Ambassadors
@Emad_A_Alghamdi
and
@zaidalyafeai
of
@arabicml2
and ASAS are hosting a sprint to ensure that Arabic isn't left behind in the development of language AI.
Join our Discord to participate!
Our community welcomes researchers, engineers, linguists, social scientists & lifelong learners from all over the world. Four months since launching, we have members from 81 countries!
Want to connect with our multi-national community? 🌏Learn more at