InPars+: Supercharging Synthetic Data Generation for IR
Krastev, M., Hamar, M., Toapanta, D., Brouwers, J., & Lei, Y. (2025). InPars+: Supercharging Synthetic Data Generation for Information Retrieval Systems. arXiv preprint arXiv:2508.13930.
CV
Education
- MSc in Artificial Intelligence, University of Amsterdam, 2023 - 2025
- Relevant Subjects: Machine Learning and Deep Learning · NLP · Computer Vision · Graph Neural Networks · Recommender Systems · Foundation Models · Reinforcement Learning
- BSc in Computer Science, University of Aberdeen, 2019 - 2023 (1st Class, GPA: 3.8/4.0)
- Relevant Subjects: Artificial Intelligence · Machine Learning · Security · Distributed Systems · Operating Systems · Modern Programming Languages
Work experience
- Teaching Assistant, University of Amsterdam, Sep 2024 - Jun 2025
- Prepared assignments and led practical sessions for the following courses: Computer Vision, Deep Learning, Information Retrieval, and Recommender Systems.
- Teaching Assistant, University of Aberdeen, Jan 2022 - Jun 2023
- Collaborated with teachers to organize and lead practical sessions for several informatics-oriented courses, including web development, operating systems, distributed systems, and discrete mathematics.
- Software Engineer Intern, Together for Success, Jun 2022 - Sep 2022
- Focus on developing a SaaS platform for building e-commerce websites, specific tasks included testing, standardization, and quality assurance. Provided extensive system documentation and requirements specification, and developed the API backend reference website.
Projects
- InPars-Plus: Supercharging Synthetic Data Generation for IR (keywords: IR, synthetic data, LLMs)
- Co-authored a method improving upon the InPars framework. We augment generation using CoT and preference fine-tuning to improve downstream performance in diverse information retrieval tasks.
- Towards Multimodal Medical Image Segmentation with Context-Prior Learning (keywords: SAM, MedAI, Segmentation)
- Co-authored a project to evaluate the transferability of SegVol, a 3D SAM-based segmentation foundation model, across different medical imaging modalities (CT and MRI). We extended SegVol’s capabilities through LoRA fine-tuning and context-priors (Gao et al., 2023), improving segmentation scores with both approaches.
- DynaLora: Dynamic Low-Rank Module Allocation (keywords: LoRA, PEFT)
- Co-authored a method for fine-tuning foundation models which dynamically allocates trainable parameters and is orthogonal to other methods, reducing the required maximum GPU memory during training.
Skills
- Programming Languages: Python, Rust, C/C++, JavaScript
- Libraries and Technologies: PyTorch, NumPy, JAX, Docker, Google Cloud, Azure
Publications
Towards Enhancing Multi-task Learning for News Recommendation
Vasilev, S., Krastev, M., & Toapanta, D. (2024). Towards Enhancing Multi-task Learning for News Recommendation.
DynaLoRA: Dynamic Low-Rank Module Allocation
Brouwers, J., Fulop, Z., Hamar, M., & Krastev, M. (2024). DynaLoRA: Dynamic Low-Rank Module Allocation.
SegEVOLution: Towards Multimodal Medical Image Segmentation with Context-Prior Learning
Z. Fülöp, S. Mihailov, M. Krastev, M. Hamar, D.A. Toapanta, S. Achlatis. (2025). SegEVOLution: Towards Multimodal Medical Image Segmentation with Context-Prior Learning.
[RE] Explaining Temporal Graph Models through an Explorer-Navigator Framework
Hamar, M., Krastev, M., Hristov, K. D., & Beglou, D. (2024). [RE] Explaining Temporal Graph Models through an Explorer-Navigator Framework.