Projects

Luth-2

Luth-2-0.8B and Luth-2-2B, French-specialized SLMs post-trained on Qwen3.5 with SFT, RL and multi-domain on-policy distillation.

• Large Language Models
• Reinforcement Learning
• Transformers
• French

Luth x LFM2

Luth-LFM2, 1.2B/700M/350M parameters models fine tuned in French in collaboration with Liquid AI.

• Large Language Models
• Transformers
• French

Luth

Luth is a state-of-the-art series of fine-tuned LLMs for French, delivering improved French instruction-following, math, and general knowledge capabilities.

• Large Language Models
• Transformers
• French

LeCarnet

LeCarnet is a 2M+ corpus of simple French stories for training small language models.

• Dataset
• Small Language Models
• French

FlashAttention-2 for sliding window attention in Triton (forward + backward).

• Triton
• Transformers
• GPU Optimization

A minimalist implementation of Paligemma 2 from scratch in PyTorch.

• Vision-Language Models
• PyTorch
• Transformers

Conditional Wasserstein Generative Adversarial Network with Gradient Penalty implementation for a Kaggle competition.

• Generative Adversarial Networks
• PyTorch

Implementation of word2vec algorithm for learning word embeddings from text corpora using skip-gram and CBOW architectures.

• NLP
• Embeddings
• PyTorch

Volleyball referee hand signal recognition in real-time or through videos.

• Computer Vision
• Sports Analytics
• PyTorch