Abdarahmane Traore
AI researcher. Currently doing a Ph.D. on multimodal large language models. Small models, big ideas, and the careful work of grounding language in pixels.
NVFP4 buys a 6x smaller rounding error than MXFP4 for a quarter of a bit
How 4-bit floats work, why MXFP4 loses small values, what NVFP4 changes about scaling, and the four-part recipe NVIDIA used to pretrain a 12B model in NVFP4 at FP8 accuracy.
Read the bit →Projects
Papers
SCOUT: Sim-to-Real Text-Based Person Retrieval by Embedding-Space Prediction over Frozen Video Features
SmolRGPT: Efficient Spatial Reasoning for Warehouse Environments with 600M Parameters
Improving Fungi Prototype Representations for Few-Shot Classification
Violence Detection in Videos using Deep Recurrent and Convolutional Neural Networks
Evaluating Deep Learning Algorithms in Pulmonary Nodule Detection
2D Bidirectional Gated Recurrent Unit Convolutional Neural Networks for End-to-End Violence Detection in Videos