Skip to content
View RohanThawait's full-sized avatar

Block or report RohanThawait

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
RohanThawait/README.md

Hi, I'm Rohan Thawait πŸ‘‹

M.Tech student at NIT Karnataka (NITK Surathkal) Β· Graduating June 2026
I train large language models from scratch and build the systems around them.


What I'm Working On

I spent the last several months going deep on LLM training β€” not using APIs, but actually training models on H100 hardware. Here's what that produced:

🧠 GRPO Reasoning Model β€” Qwen2.5-7B

Implemented the full reasoning training pipeline from the DeepSeek-R1 paper.
SFT cold start on 27K math CoT examples β†’ GRPO reinforcement learning with verifiable reward functions.
The most interesting finding: diagnosed reward saturation from training curves β€” frac_reward_zero_std averaged 0.63, meaning 63% of batches produced near-zero gradient signal. Replicates the curriculum selection problem DeepSeek documented.

πŸ“ Long-Context RoPE Scaling β€” Mistral-7B

Extended Mistral-7B context from 8K β†’ 32K tokens using YaRN RoPE scaling + LongLoRA finetuning.
Evaluated with needle-in-a-haystack tests across context lengths.


Technical Stack

LLM Training
GRPO SFT Full Finetuning LoRA Reward Function Design BF16 FlashAttention-2 Gradient Checkpointing

Frameworks
PyTorch HuggingFace TRL Transformers DeepSpeed lm-evaluation-harness Weights & Biases

Inference & Evaluation
Speculative Decoding YaRN RoPE Scaling GSM8K MATH ARC benchmarks

Application Layer
LangChain LangGraph FastAPI Streamlit RAG FAISS

Infrastructure
H100 NVL CUDA Linux Docker Git HuggingFace Hub


Other Projects

Project What it does
VoicePaper Converts research papers into podcast audio β€” PyMuPDF β†’ GPT-4 β†’ ElevenLabs
LawyerLens RAG pipeline for legal document Q&A β€” FAISS + OpenAI embeddings
FinancialAgent LLM agent with conversational SQL, data viz, Plaid API integration

Background

  • πŸŽ“ M.Tech Computational and Data Science β€” NIT Karnataka (Surathkal) (GATE CS 2024)
  • πŸŽ“ B.Tech Computer Science Engineering β€” SAGE University
  • πŸ’‘ 500+ LeetCode problems Β· Kaggle Contributor Β· Python Gold @ HackerRank

Let's Talk

I'm open to LLM Engineer, ML Engineer, and AI Engineer roles β€” India and remote international.

πŸ“§ thawaitrohan@gmail.com
πŸ€— huggingface.co/thawait
πŸ’Ό LinkedIn]

Pinned Loading

  1. qwen2.5-7b-math-reasoning-grpo qwen2.5-7b-math-reasoning-grpo Public

    This project implements the reasoning training pipeline introduced in the DeepSeek-R1 paper, applying Group Relative Policy Optimization (GRPO) to teach a 7B language model to reason step-by-step t…

    Python

  2. LawyerLens LawyerLens Public

    An AI legal assistant that analyzes and validates your documents by comparing them against a curated knowledge base of Indian law.

    Python

  3. CodeQuest CodeQuest Public

    An AI-powered Slack bot that accelerates software engineer onboarding by answering questions about documentation and explaining code using LangChain and Google Gemini.

    Python

  4. FinancialAgent FinancialAgent Public

    An advanced AI financial agent built with LangChain and Google Gemini. Features conversational SQL, on-demand data visualization, and real-time transaction syncing via the Plaid API.

    Python

  5. VoicePaper VoicePaper Public

    A full-stack, containerized application that transforms online articles and research papers into on-demand podcasts using FastAPI, Streamlit, PostgreSQL, and Docker.

    Python