Hi, I'm Hannan Mahadik 👋

“Research is about pushing boundaries, and mentoring is about helping others find theirs.”

— David Clutterbuck

I'm currently pursuing a PhD at the ELLIS Institute Tübingen, where I focus on research in Large Language Models (LLMs), deep learning, and model compression. I am part of the OpenEuroLLM project (read more here). I also enjoy mentoring students and sharing knowledge with the community.

• Post-Trainining (SFT + DPO) and Evaluation (static + LLM-as-a-judge) for Language Models
• Model Compression (Pruning) and Distillation (both via synth data gen and traditional logits distillation)
• Exploring ways to expand my master thesis (Fairness in Recommendations using Graph Neural Networks)

Open to collaboration and discussion on any of these topics! 😃

• Deep Learning, including Generative Models like VAEs, Diffusion models, GANs etc.
• LLMs: Post-training, Evaluation, Synthetic Data Generation, Pretraining (not as much)
• Model Compression (Pruning) & Knowledge Distillation
• Graph Neural Networks (GNNs)
• Teaching & Mentoring

• GNNs-FAME: My master thesis on mitigation of biases in Graph Neural Networks for fairer recommendations
• Whittle: Python library to compress LitGPT models for resource efficient inference
• arena-hard-auto-private: Data analysis on language models judged by an LLM using Arena-Hard
• synetune-annealing-experiments-oellm: Using Synetune to find data weights for annealing experiments
• I also have my assignments from university for courses like Machine Learning, Deep Learning, Generative Models, etc. uploaded here.

• "whittle: A Library for Sub-Network Extraction from Large Language Models" AutoML 2026 (ABCD Track, Oral). Best Paper Award in the ABCD Track. OpenReview

  We introduce whittle, an open-source library for extracting structured sub-networks from pretrained model checkpoints, providing a principled approach to initializing small language models and to model compression. Among its components, whittle exposes a sub-network extraction API alongside pretraining and knowledge-distillation pipelines, and is designed so that the extraction primitive can be driven by external search procedures over arbitrary user-defined sub-network search spaces. Using whittle, we identify sub-networks of Pythia-6.9B and Pythia-12B that match the validation perplexity of comparable Pythia baselines while requiring up to 5.16× and 1.26× fewer floating point operations at 10B and 100B tokens, respectively. We further demonstrate the generality of the extraction API on Qwen3-8B, where the same primitive supports three external search methods: random search, evolutionary search, and Bayesian optimization with the Tree-structured Parzen Estimator. We release the code of whittle at https://github.com/whittle-org/whittle, and the code for the experiments at https://github.com/whittle-org/whittle-paper .

• "GNN's FAME: Fairness-Aware MEssages for Graph Neural Networks" Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization (UMAP '25), 2025.
  DOI: 10.1145/3699682.3728324

  Graph Neural Networks (GNNs) have shown success in various domains but often inherit societal biases from training data, limiting their real-world applications. Historical data can contain patterns of discrimination related to sensitive attributes like age or gender. GNNs can even amplify these biases due to their topology and message-passing mechanism, where nodes with similar sensitive attributes tend to connect more frequently. While many studies have addressed algorithmic fairness in machine learning through pre-processing and post-processing techniques, few have focused on bias mitigation within the GNN training process. In this paper, we propose FAME (Fairness-Aware MEssages), an in-processing bias mitigation technique that modifies the GNN training’s message-passing algorithm to promote fairness. By incorporating a bias correction term, the FAME layer adjusts messages based on the difference between the sensitive attributes of connected nodes. FAME is compatible with Graph Convolutional Networks, and a variant called A-FAME is designed for attention-based GNNs. Experiments conducted on three datasets evaluate the effectiveness of our approach against three classes of algorithms and six models, considering two notions of algorithmic fairness. Results show that the proposed approaches produce accurate and fair node classifications. These results provide a strong foundation for further exploration and validation of this methodology.

• Open LLM Research
• Model Compression and Knowledge Distillation
• Teaching and mentoring
• I enjoy collaborating and exchanging ideas on AI-related topics; always open to new research collaborations.

• I played U-19 and U-16 Cricket for the Kuwait National Team.
• I have been an international student mentor at OvGU, Magdeburg, for over three years, and have mentored ~500 students.

I speak the following languages fluently:

Message me on: