Language Breakdown
Lines of code distribution across 15 owned repositories
T-Shaped Developer
T-shapedDeep in HTML with broad versatility
Collaboration Network
Global Impact visualization
Repos
15
PRs
0
Growth
+18%
Top Collaborators
No collaborator data yet.
Coding Streak
Contribution activity over the past year
Top Repositories
Reproducing Lora Without Regret blog from scratch + Researching the ratio between LoRA and FullFT learning rates
training models with different optimizers using NumPy only. Featuring SGD, Adam, Adagrad, NAG, RMSProp, and Momentum. This repo also includes a benchmark against Pytorch developed optims.
Training tiny models from scratch using NumPy in code and linear algebra on a piece of paper.
The Lightest Vision Transformer (ViT) trained from scratch out there to achieve 93.37 ± 0.07%” top-1 accuracy on CIFAR-10 within just 50 epochs.
Replication of "Adversarial Examples Are Not Bugs, They Are Superposition" paper.
An empirical study that builds on Merrill et al. theoretical construction to solve majority. A closer look into transformers reaction to training dynamics and data corruption.
A CLI that converts CSV files and folders with image data to a JSON file with COCO annotations and builds a custom dataset off of it ready to train for object detection tasks.
A clean and modular pipeline for preprocessing the Food-101 dataset using both folder-based and CSV-based workflows.
Interactive visualization simulating how Gamma and Beta control ICS (Internal Covariate Shift) for faster convergence in BatchNorm.
PISSA-Without-regret is Pissa used with Lora-Without-Regret blog optimal configuration. We provide comparisons to LoRA and FullFT using the same regime.
Open Source Impact
Contributions to external projects
