Machine Learning Engineer @ LightOn
- 🚀 LightOnOCR, a family of efficient 1B end-to-end OCR VLMs — v2 achieves SOTA on OlmOCR-Bench while being 9× smaller and up to 5× faster than competing approaches
- 🏗️ ModernBERT, contributed to architecture design, training and eval (ACL 2025)
- 🌐 ArabicWeb24, a 39B token Arabic corpus for LLM training
- 🛠️ vit.cpp, a lightweight C++ inference engine for Vision Transformers using GGML
- 💬 Interested in Vision Language Models, Vision Transformers, LLM Pre-training, State-Space Models, Optimization, Code Generation, Efficient Inference, Quantization, GPU Kernels, Distributed Training, RL
- 🎓 Engineering degree in maths and machine learning from École Centrale de Lyon
- 📫 Reach me: taghadouinisaid@gmail.com
