FeeL: Making Multilingual LMs Better, One Feedback Loop at a Time

riddhib Riddhi Bhagwat

borgr Leshem Choshen

MIT | HuggingFace | IBM | Cohere | KAIST | BIU

Welcome to FeeL

Have you ever tried chatting with an open source language model in a language other than English only to get a response that barely makes sense?? You're not alone.

Most modern AI improvements only prioritize widely spoken languages, while other languages trail behind. But what if we could build better AI for all languages - by users for users?

FeeL - Feedback Loop – is a first-of-its-kind, open platform for improving language models through use and feedback. Whether your language is Dutch, Arabic, Korean, Hebrew, Hindi, or otherwise, you can use the model and it will learn to fit your language preferences through FeeL.

The FeeL Story 📖

Language models learn and evolve based on feedback through their interactions. However, the resulting behavior for closed models (e.g., ChatGPT,Gemini, Claude, etc.) is ultimately shaped by the companies controlling them, while open models are often shared once and left unchanged – neither ever truly adapting to user needs. These models also do not account for different cultures associated with different language communities, and there is limited open data to improve models with.

The fundamental issue here becomes apparent:

• A lack of high-quality open-source datasets
• No method to continuously improve open models
• Community has little control over model behavior

Even with mainstream AI companies making large improvements in multilingual capabilities, the data and improvement pipelines are unable to be used outside of the closed doors of companies.

Enter FeeL 🎉

Each of us on the FeeL development team came with our own set of similar experiences with open source language models and, with a shared love for advancing technology for the greater good, FeeL was born.

As we aspire to expand technology on a global scale, we must also cater to the cultures, values, and, of course, languages of global communities.

FeeL is about flipping the script - letting real users shape how open models evolve in their own languages. By chatting with the models, correcting their mistakes, and giving feedback, speakers of different languages can directly influence how well AI understands and responds in their native tongue for better and more cognizant AI.

The more diverse the feedback, the better these models become for everyone.

How does FeeL work?

FeeL is an open source platform on the HuggingFace space where users can select their preferred language and chat with the model.

As a response is generated, you can do any of the following:

If the model response looks great:

• Give it a thumbs up 👍 - Quick feedback to let us know the model got it right!

If the model response needs work:

• Regenerate 🔄 - See if another try gives a better result.
• Give it a thumbs down 👎- Quick feedback to let us know the model needs work!
• Edit the response ✏️ - Rewrite it to sound more natural.

Conversation data, if you choose to submit it, transfers from the user side directly into the dataset, which is made open source for fostering multilingual AI development. Your feedback feeds directly into an RLHF (Reinforcement Learning from Human Feedback) pipeline to fine-tune and improve the model in real time.

Join the Loop!

FeeL is a project by users for users (like you!). If you're passionate about bringing smarter models to your language, join FeeL, chat with the models, give feedback, and watch them improve - one response at a time. If you have a different community that models should support better, feel free to duplicate and adapt, or reach out! We love your help in making AI better for everyone. 🚀