Omneity Labs

[Omar Kamali @OmarKamali]

Sawalni Live mode - beta preview 1

@laytoun wahed early preview dial TTS, mazal in training ✌️

[Omar Kamali @OmarKamali]

@Rainmaker1973 I built Sawalni from scratch, a Moroccan AI free from OpenAI or Google influence 🇲🇦

Discover Morocco’s culture and languages with Sawalni and make sure to follow @SawalniBot to support the project ✌️

https://t.co/xMtlkzrpHu

[Omar Kamali @OmarKamali]

I got Qwen to count R in Strawberry correctly!🚨

LLMs are good at writing. They’re worse at words, and this is one of the main blockers to true multilinguality.

[Omneity Labs @OmneityLabs]

Visit the LID benchmark here

https://t.co/v2fXbFtUlk

[Omneity Labs @OmneityLabs]

🌐 LID Benchmark update

• 10 Regional Leaderboards
• 17 LID models (+7 new, incl. non-fastText based)
• 449 languages in total (200+ additional)
• Fixed: F1 macro reporting error
• Normalized language codes for more accurate results

[Omneity Labs @OmneityLabs]

Babelcode is a library for normalizing language codes sanely:

- ISO 639-1 (en)
- ISO 639-3 (eng)
- BCP-47 (zh-Hans)
- NLLB-style (eng_Latn)
- WikiPron filenames (wp_eng_latn_us)
- CHILDES corpus names (EnglishNA)
- or plain natural language (German)

https://t.co/BSSd05Wctb

[Omneity Labs @OmneityLabs]

7+ standards, one library to rule them all.

It's one of those problems that keeps breaking multilingual pipelines: language codes.

Language names change, codes change, formats differ, and macrolanguages make everything more interesting.

So we built and released Babelcode.

[Omar Kamali @OmarKamali]

Language ID Leaderboard is live! 🌐

• 10 LID models including our very own Gherbal leading the pack 😎.

• 8 benchmark datasets including the recent BOUQuET dataset from Meta AI.

• 380 languages and various regional lenses.

[Omar Kamali @OmarKamali]

My custom LLM architecture in action (real data captured from executing it)

[Omar Kamali @OmarKamali]

I just might have cracked tokenizer-free LLMs. No vocab, no softmax.

[Omar Kamali @OmarKamali]

Teaching an LLM a new language is so effortful but can be rewarding and high impact.

In my recent interview with @FeatherlessAI (thank you for having me!), I recount the story of training @SawalniBot, the challenges, and many learnings I wish I knew before.

Enjoy the read!

[Omneity Labs @OmneityLabs]

We are live! 🎙️