pytorch

7.67× LoRA / 8.35× Full FT speedup for Qwen3.5 (0.8B–27B) on NVIDIA DGX Spark — wall-clock parity with rented H100. Lossless within BF16. Three-command interactive wizard handles model picker, data validator, training, and merge.

Complete setup guide for a 2-node NVIDIA DGX Spark cluster — distributed training, CUDA inference with EXO, NCCL tuning for Grace Blackwell, NVMe-TCP shared storage, and 200 Gb/s direct fabric networking.

Advance information retrieval system that combines advanced indexing, machine learning, and personalized search to enhance academic research and document discovery.

Tracking exactly what happens to the internal "circuitry" (induction heads) of a 2-layer attention-only Transformer when forced to undergo domain adaptation from prose to structured Python code.

Detecting fake news articles by analyzing patterns in writing.

Official codebase for RetFiner: A Vision-Language Refinement Scheme for Retinal Foundation Models

A voice assistant with local LLM as a backend

Different post-training techniques for LLMs, including: SFT, DPO and Online RL

An end-to-end tennis video analysis system leveraging Ultralytics YOLOv8 for player and ball detection, multi-object tracking across frames, and a custom CNN for court keypoint extraction, providing data-driven insights into player movement, ball trajectories, and match dynamics using advanced computer vision and deep learning techniques.

Library for handling atomistic graph datasets focusing on transformer-based implementations, with utilities for training various models, experimenting with different pre-training tasks, and a suite of pre-trained models with huggingface integrations

Omega-Forge: unofficial code to finetune or train VGGT-Omega (VGGT-Ω) on your own dataset, for use while official training code is not yet available. Additionally includes a standalone zero-GPU 3D reconstruction visualizer.