Open Science Community
Research notes, technical essays, and personal stories from the Cohere Labs Community
Research notes, stories, and ideas from people shaping AI together—transparent, collaborative, and community-led.
How a Luganda word-review task became a startup, a Gold Award, and a lesson about what really keeps gates closed for African builders, and what opens them.
We trained and released Sparse Autoencoders on all four Tiny Aya regional variants and asked what an SAE sees inside a model built for 70+ languages: a representational-density gap, a mostly-shared feature basis, and an identity-versus-quality null.
We adapt Kyutai's Moshi architecture for Turkish↔Hindi speech-to-speech translation by replacing the audio-pretrained backbone with Cohere's multilingual TinyAya, and share what worked, what broke, and what we learned.
Sinhala NLP has often been framed as a data-scarcity problem, but the next major step is better evaluation: benchmarks that test real Sinhala, local context, code-mixing, and everyday use.
RLVR fine-tuning raises accuracy on medical VQA benchmarks while quietly degrading visual grounding: a new counterfactual evaluation framework identify the gap.
