Master Thesis: Evaluating synthetic data generation for delta learning
AI Sweden
Summary
As large language models (LLMs) scale, gathering high-quality human preference data for instruction tuning becomes a critical bottleneck. In OpenEuroLLM, we are using natural, synthetic, and translated data. Recent experiments indicate that models trained on translated data often suffer from "translationese" artifacts, masking native fluency. Delta learning has recently been put forward as an approach for steering models towards specific qualities: Using DPO to train models on datapoints that contrast in a single qualitative dimension we can steer model behaviour in that dimension. One such dimension is fluency: Can we synthesize documents that differ only in the fluency/translationese dimension, so that we can steer models trained on less fluent data towards being more fluent, while retaining other capabilities?
Job description
Project Background and Problem Statement
AI Sweden and OpenEuroLLM are currently training Large Language Models for all European Languages. Machine translated or synthetic data will be necessary to cover all languages, at all stages of training, but machine translated data will introduce some amount of translationese or poor fluency. Delta learning offers a potential way to mitigate this: By isolating the unwanted quality (lack of fluency / translationese) and generating or identifying documents that differ in that quality alone. The core question is:
- Can Delta Learning be employed in this way? How to source data such that it differs only in the desired quality? And what (if any) are the adverse effects on other tasks, if this is applied? (As measured on MMLU ProX or Belebele).
Outline
The goal of this project is to rigorously evaluate DPO data synthesis strategies on Prelude 9B.
- Literature study: Review DPO methodologies, synthetic data generation pipelines, and LLM evaluation benchmarks focusing on translationese artifacts.
- Implementation: Construct controlled DPO training runs using Prelude 9B. Compare a baseline before fluency mitigation against runs after fluency mitigation, comparing the effect of data choices, training hyperparameters, et.c. on chosen benchmarks.
- Evaluation: Evaluate fluency, safety, and adequacy using JudgeArena and standard multilingual benchmarks to measure the Elo rating improvements.
Who we’re looking for
We are seeking curious, self-driven MSc students eager to work at the frontier of open-weight European AI research (LLMs). You thrive on empirical discovery, design rigorous experiments, and let data challenge your assumptions.
- Ongoing Master’s studies in Computer Science, Data Science, Machine Learning, Engineering Physics, or a related quantitative field.
- Proficiency in Python and hands-on experience with modern deep learning frameworks (PyTorch, Hugging Face ecosystem).
- Familiarity with LLM post-training alignment (e.g., SFT, DPO, RLHF/RLVR) or context-extension, alongside comfort running distributed GPU training in Linux/HPC environments.
Requirements
- Ongoing Master’s studies in Computer Science, Data Science, Machine Learning, Engineering Physics, or a related quantitative field.
- Proficiency in Python and hands-on experience with modern deep learning frameworks (PyTorch, Hugging Face ecosystem).
- Familiarity with LLM post-training alignment (e.g., SFT, DPO, RLHF/RLVR) or context-extension, alongside comfort running distributed GPU training in Linux/HPC environments.
Company overview
AI Sweden is Sweden's national center for applied artificial intelligence, established in 2019 with a mission to accelerate the adoption and implementation of AI technologies for societal benefit, economic competitiveness, and national advancement. As a collaborative research institute, it brings together over 170 partners from diverse sectors including government agencies, universities, research institutions, startups, and global corporations. The organization operates as a catalyst for AI innovation, focusing on practical applications across healthcare, energy systems, municipalities, and various industries. With a core team of over 80 professionals comprising AI experts, data scientists, research engineers, linguists, mathematicians, policy experts, and change leaders, AI Sweden facilitates knowledge sharing, talent development, and infrastructure building. The institute operates on a not-for-profit basis, funded by Sweden's innovation agency Vinnova, the Swedish Agency for Economic and Regional Growth, European Regional Development Fund, Västra Götalandsregionen, and partner contributions. Through its AI Labs, talent programs, workshops, and collaborative projects, AI Sweden addresses critical societal challenges while bridging the gap between cutting-edge research and real-world implementation, positioning Sweden as a leader in responsible and impactful AI development.