Yuvraj Singh

Yuvraj Singh

LLM · RL · Distributed Systems

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

Posts

Future Blog Post

less than 1 minute read

Published:

This post will show up by default. To disable scheduling of future posts, edit config.yml and set future: false. Read more

Blog Post number 4

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Headings are cool

You can have many headings

Aren’t headings cool?

Read more

Blog Post number 3

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Headings are cool

You can have many headings

Aren’t headings cool?

Read more

Blog Post number 2

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Headings are cool

You can have many headings

Aren’t headings cool?

Read more

Blog Post number 1

less than 1 minute read

Published:

This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.

Headings are cool

You can have many headings

Aren’t headings cool?

Read more

datasets

QnA Irrigation Diseases Dataset

This dataset contains a comprehensive Question & Answer collection focused on water management technologies, irrigation systems, and related agricultural practices for sustainable farming. The dataset is derived from technical documentation and research publications related to water management in ag
Read more

QnA Plant Diseases Dataset

This dataset contains a comprehensive Question & Answer collection focused on plant diseases, their management, treatment protocols, and diagnostic techniques. The dataset provides detailed information about various plant pathologies, fungicide applications, and disease identification methods for ag
Read more

QnA Soil Diseases Dataset

This dataset contains a comprehensive Question & Answer collection focused on soil management, soil health, organic farming practices, and soil-related agricultural techniques. The dataset is derived from technical guides and documentation related to sustainable soil management practices, with empha
Read more

models

ViT

Published:

ViT-B/16 from scratch on a 3-class Food-101 subset. Train loss 1.20 / test loss 1.52. Read more

GPT

Published:

Decoder-only transformer trained on TinyShakespeare, replicating the original OpenAI GPT architecture from scratch. Read more

BERT

Published:

Bidirectional encoder pre-trained with masked language modelling on the Cornell Movie Dialogs corpus. Read more

CycleGANs

Published:

Cycle-consistent unpaired image translation on Cityscapes — two generators, two discriminators, cycle + identity losses. Read more

Differential Transformer

Published:

Differential attention replicated from scratch — two attention maps subtracted to cancel noise. Trained on TinyShakespeare on A100. Read more

Encoder-Decoder

Published:

LSTM-based Seq2Seq encoder-decoder for German→English translation. Train/val loss ~1.38 in 10 epochs. Read more

GRU

Published:

GRU from scratch. 16 hidden units, 50 epochs. Train loss 0.51 / val loss 0.48. Read more

RNNs

Published:

Vanilla RNN from scratch. 16 neurons, 50 epochs. Train loss 0.51 / val loss 0.50. Read more

Transformer

Published:

Encoder-decoder transformer for English→Hindi translation on Samanantar (~25M params). Published on HuggingFace. Read more

Mixtral

Published:

Sparse MoE transformer replicated from scratch on TinyShakespeare. Train loss 2.04 / val loss 2.09 in 1,000 steps on T4. Read more

DPO

Published:

Direct Preference Optimization applied to Qwen0.5B-Instruct on UltraFeedback. Train loss 0.67 in 3,000 iterations. Read more

SimplePO

Published:

Reference-free preference optimization (SimplePO) on OPT-330M. Batch size 128, lr=2e-5, beta=2 on UltraFeedback. Read more

LoRA

Published:

Low-rank adaptation implemented from scratch in PyTorch. Train/val loss ~3.5 in 1,000 steps on A100. Read more

ORPO

Published:

Odds Ratio Preference Optimization on OPT-330M. Reference-free alignment reaching train loss 1.70 in 3,000 iterations. Read more

Gemma

Published:

Google’s Gemma architecture replicated from scratch — multi-query attention and GeGLU activations on TinyShakespeare. Read more

Llama

Published:

Decoder-only Llama replicated from scratch with RoPE, SwiGLU, RMSNorm and GQA. Read more

CLiP

Published:

Contrastive vision-language model trained on Flickr8K. Train loss 1.3 / val loss 2.2 in 30 epochs on T4. Read more

DDP

Published:

Llama trained with PyTorch DistributedDataParallel (torchrun). Val loss 1.1 in 8,000 iterations on TinyShakespeare. Read more

Llava

Published:

Visual instruction tuning replicated from scratch on Flickr8K. Train loss 0.23 / val loss 0.22 in 5 epochs on T4. Read more

Seq2Seq

Published:

GRU-based Seq2Seq with both Bahdanau and Luong attention from scratch. 128 hidden units, 50 epochs. Read more

Whisper

Published:

Whisper ASR from scratch — CNN on 80-channel mel spectrograms + 6-layer transformer decoder. Trained on GigaSpeech. Read more

LSTM

Published:

LSTM from scratch (~128K params). 128 hidden units, 50 epochs. Train loss 0.49 / val loss 0.48. Read more

Gemma3

Published:

90M-parameter Gemma 3 with local sliding-window attention (128-token blocks). Val loss 1.77 in 25k steps on TinyStories. Read more

Llama4

Published:

1.2B-parameter MoE (32×12M experts, top-1 routing) trained on TinyStories. Val loss 1.70 in 20k steps on Kaggle P100. Read more

Moonshine

Published:

Compact transformer ASR (288-dim, 6 heads) trained on GigaSpeech for 1,500 steps. Notes on overfitting at ~25 hours. Read more

PaliGemma

Published:

Google’s PaliGemma VLM (SigLIP + Gemma) replicated from scratch on Flickr8K. Read more

Pix2Pix

Published:

Conditional GAN for paired image-to-image translation (aerial→map) replicated from scratch. PatchGAN discriminator. Read more

SigLip

Published:

Sigmoid-loss vision-language pretraining replicated from scratch on Flickr8K — avoids global softmax normalisation. Read more

TTS

Published:

Tacotron-style transformer TTS from scratch — 512-dim phoneme encoder, mel spectrogram decoder, 16kHz on GigaSpeech. Read more

VAE

Published:

VAE on CelebA (128×128). 4-layer conv encoder, 32D latent, ConvTranspose decoder. Reconstruction + KL loss over 200 epochs. Read more

WGANs

Published:

Wasserstein GAN and WGAN-GP implemented from scratch on MNIST — gradient penalty for stable training. Read more

Kimi-K2

Published:

DeepSeekV3-inspired MoE with latent attention trained with Muon optimizer. Pre-trained weights on HuggingFace. Read more

CGANs

Published:

Conditional GAN on MNIST — class-conditioned 64×64 digit generation. 30 epochs, BCE loss, TensorBoard logging. Read more

CLAP

Published:

Contrastive Language-Audio Pretraining from scratch on GigaSpeech. 768D text / 2048D audio → 1024D shared space. Read more

DCGANs

Published:

Deep Convolutional GAN trained on CelebA and CIFAR-10. ~7,800 steps (CelebA) and ~11,700 steps (CIFAR-10). Read more

DeepSeekV3

Published:

16×4 MoE with Multi-head Latent Attention and auxiliary-free load balancing, trained on TinyStories on Kaggle P100. Read more

portfolio

projects

Movies Review System Spoiler-Free Sentiment-Analysis based Movies Review System)

Published:

Introducing the Movie Review System, where AI meets movie magic to revolutionize how viewers experience films. This project goal is to provide an interface for spoiler-free reviews and sentiment analysis, enhancing the viewing journey. With advanced models like Voting Classifier and Bi-LSTMs powered by Keras and TensorFlow, we achieve impressive metrics—a 91% accuracy, 91% precision, and 90% recall that understands and enhances the users’ with the respective movies’ plot from a bird’s eye view. Read more

MoviesMania (Geek-o-thon) A Reverse Search based Movies Recommendation System

Published:

Step into the future of entertainment discovery with MoviesMania. The rpoduct aims to simplify your search for the perfect movie or web series. Using various AI/ML techniques and elements, we analyze uploaded video clips to predict movie titles and recommend similar content with an impressive accuracy. Experience flavoured recommendations tailored to your tastes, powered by Keras, Flask, and advanced face recognition algorithms with a full-fledged movie recommendation system. Read more

PlogPayouts AI-driven Plogging System

Published:

Transform your daily jog into a mission for a cleaner world with PlogPayouts. Our innovative website + app rewards you for collecting litter, promoting fitness and environmental cleanliness. Utilizing AI for trash categorization and optimized routes, and fostering community through shared stories, PlogPayouts turns every step into a step towards a greener, more inclusive society. Join us and make a difference today! Read more

Insight-Ed (HackNITR 5.0) EdTech Platform for Student and Teacher

Published:

Imagine an online classroom where teachers instantly know when and why students lose focus. Our AI-powered solution bridges the knowledge gap by detecting student emotions and attentiveness, highlighting problem areas, makes the teacher aware of each student’s progress. With features like reverse video search, dynamic questionnaires, and advanced Q&A bots, we transform the learning experience, making it more interactive and insightful. Redefining online education with our model, ensuring every student gets the attention they need, right when they need it. Read more

FarmGenie (GeoHack 2024) Empowering farmers with real-time insights and expert guidance via AI-driven space

Published:

Our platform utilizes LLMs and a Mixture of Expert (MoE) approaches to provide precise guidance on soil management, plant disease identification, and irrigation techniques. Built as a scalable web application with a Next.js frontend and backend, and supported by a Redis queue and multiple worker nodes, FarmGenie ensures robust performance. The system’s multilingual support, interactive community forum, and up-to-date knowledge base facilitate seamless, expert-driven assistance for both new and experienced farmers. Read more

NeatRL Playground AI Games Showcase powered by Reinforcement Learning

Published:

Beautiful, interactive website showcasing AI-powered games with reinforcement learning agents. Features Pong AI with Deep Q-Learning, real-time WebSocket communication, and smooth animations. Deployed on Vercel (frontend) and Render (game server) with production-ready health checks and headless mode. Read more

NeatRL Deep Reinforcement Learning Algorithms Library

Published:

Comprehensive implementations of deep RL algorithms including DQN, A2C, PPO, DDPG, TD3, and SAC. Features one-file implementations, experiment tracking with W&B, automatic video recording, and support for Gymnasium environments. Main NeatRL library provides high-quality training utilities with focus on simplicity and performance. Read more

publications

rl

Q-Learning

Published:

Tabular Q-Learning and Value Iteration implemented from scratch as educational notebooks. Read more

DQN Flappy

Published:

DQN agent trained on Flappy Bird using pixel observations, experience replay, and epsilon-greedy exploration. Read more

VizDoom RL

Published:

DQN agent trained on VizDoom Basic via Gymnasium wrapper, with grayscale preprocessing, replay buffer, and W&B logging. Read more

GRPO

Published:

Group Relative Policy Optimization — DeepSeek-R1’s RL training objective implemented from scratch. Read more

A2C (A2C)

Published:

Implementation of A2C reinforcement learning algorithm Read more

DDPG

Published:

Implementation of DDPG reinforcement learning algorithm Read more

DQN Lunar

Published:

Implementation of DQN-Lunar reinforcement learning algorithm Read more

DQN Taxi

Published:

Implementation of DQN-Taxi reinforcement learning algorithm Read more

DQN Atari

Published:

Implementation of DQN-atari reinforcement learning algorithm Read more

DQN

Published:

Implementation of DQN reinforcement learning algorithm Read more

Duel DQN

Published:

Implementation of Duel-DQN reinforcement learning algorithm Read more

Frozen Lake

Published:

Implementation of Frozen-Lake reinforcement learning algorithm Read more

MARL

Published:

Implementation of MARL reinforcement learning algorithm Read more

IPPO

Published:

Implementation of IPPO reinforcement learning algorithm Read more

MAPPO

Published:

Implementation of MAPPO reinforcement learning algorithm Read more

Self Play

Published:

Implementation of Self Play reinforcement learning algorithm Read more

PPO

Published:

Implementation of PPO reinforcement learning algorithm Read more

Atari

Published:

Implementation of Atari reinforcement learning algorithm Read more

REINFORCE

Published:

Implementation of REINFORCE reinforcement learning algorithm Read more

RND

Published:

Implementation of RND reinforcement learning algorithm Read more

SAC

Published:

Implementation of SAC reinforcement learning algorithm Read more

TD3

Published:

Implementation of TD3 reinforcement learning algorithm Read more

smolhub

Smol Mixtral

Published:

A PyTorch implementation of a Mixtral inspired transformer model with Mixture of Experts (MoE), designed for text generation and understanding tasks. This model is built on the Mixtral architecture wi…
Read more

Smol Transformer

Published:

A compact implementation of an Encoder-Decoder Transformer for sequence-to-sequence translation tasks. This project implements a translation model from English to Hindi using the Samanantar dataset.
Read more

Story Kimi

Published:

A PyTorch implementation of a DeepSeek V3 inspired transformer model with Mixture of Experts (MoE), Latent Attention, and other advanced features.
Read more

Story Llama

Published:

So, I trained a Llama a 88M architecture I coded from ground up to build a small instruct model, going through the below-mentioned stages from scratch.
Read more

Story Mixtral

Published:

A PyTorch implementation of a Mixtral inspired transformer model with Mixture of Experts (MoE), Flash Attention, and other advanced features.
Read more

Smol Llama

Published:

So, I trained a Llama a 130M architecture I coded from ground up to build a small instruct model, going through the below-mentioned stages from scratch.
Read more

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Heading 1

Heading 2

Heading 3

Read more

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.

Heading 1

Heading 2

Heading 3

Read more