Giving Back

===

• RLFT
  A lightweight research-oriented project built on nanoGPT (Andrej Karpathy) to study reinforcement learning–based fine-tuning on a controlled toy setting.
  The model is trained on a digit arithmetic task, where:
  • Addition is learned via supervised fine-tuning.
  • Subtraction is improved using Group Relative Policy Optimization (GRPO).

  The project further analyzes token embedding geometry before and after GRPO-based fine-tuning to investigate whether emergent structure or behavior arises in the latent space, particularly from supervised learning on addition alone.

  Code and experiments: https://github.com/subratpp/rlft

• F1Tenth Tried to setup F1Tenth car from scratch till simple pure pursuit algorithm. https://github.com/subratpp/f1tenth