SmolLM3-RAG: A High-Performance, Low-Cost RAG System

Reasoning-style fine-tuning of an instruction LLM using LoRA vs QLoRA, analyzing the accuracy–memory trade-off on CommonsenseQA with real GPU constraints.

Sherlock-LLM investigates persona alignment in large language models through parameter-efficient fine-tuning. The project studies how lightweight adaptation methods can enforce consistent persona traits in Qwen-2.5 models without full model retraining, making the approach scalable and practical for research use.

This project presents a medical question–answering language model built by fine-tuning Google Gemma-2-2B-IT using LoRA (Low-Rank Adaptation) 🧠⚕️. The primary objective is to adapt a general-purpose large language model to the healthcare domain in a parameter-efficient, reproducible, and resource-aware manner.

[ELAMI-MICCAI 2025] Official implementation of "Pixels Under Pressure: Exploring Fine-Tuning Paradigms for Foundation Models in High-Resolution Medical Imaging"

A mixture-of-LoRA-experts system that dynamically routes instruction queries to specialized adapters based on task complexity estimation. Trains multiple LoRA modules on complexity-stratified Alpaca data with a neural router achieving 96.9% routing accuracy at 3.97ms latency.

Geometry dynamics and instability in federated LoRA under heterogeneous data distributions (FedGeoX)

Spam Email Classification using LoRA Fine-tuned Transformers: High-performance spam email classification using LoRA-adapted transformer models (ELECTRA, RoBERTa). Achieves 99.4%+ accuracy with parameter-efficient fine-tuning on 83K+ emails.