Postdoctoral Researcher in Language Model Interpretability and Digital Humanities

You will lead the project's computational work, co-supervised by the PI and the project partner Dr. Juan Luis Gastaldi (Sorbonne University and ETH Zürich). Your tasks include:

preparing and validating the historical corpora (OCR quality, genre and date metadata) with the project's research assistant

fine-tuning open-weight language models on period texts (continued pretraining with parameter-efficient methods such as LoRA)

building the attribution pipeline, with token-level SHAP checked against probing and causal-intervention methods

running systematic model tests and controlled prompting experiments with the team's literary historians

leading the technical side of the benchmark release (documentation, baselines, public leaderboard)

leading or co-authoring papers at NLP and machine-learning venues and co-authoring the project's humanities publications

The position supports your own research profile and further academic qualification.

The Digital Humanities Laboratories (DHLabs) at the University of Basel are an interdisciplinary unit for research, teaching and infrastructure on technological and digital worlds in the humanities and social sciences. Within the DHLabs, contralab, the research group of Prof. Dr. Ranjodh Singh Dhaliwal, brings together machine learning and NLP, literary history and theory, media studies and STS, and the philosophy of language. For the four-year project “A Philology of Probabilistic Texts: The Problem of Interpretation in Artificial Intelligence,” funded by the Swiss National Science Foundation, contralab is looking for a postdoctoral researcher (80%) as of 1 March 2027 or by arrangement. The project asks what large language models do to meaning when they read historical texts: it maps how eighteenth-century concepts shift across genres and decades in large historical corpora, examines how fine-tuned open-weight models represent those concepts, and releases an annotated corpus of model divergences, an open benchmark for diachronic semantic understanding, and a taxonomy of the ways models misread historical texts.

• A doctorate (by the time of appointment) in computer science, NLP or computational linguistics, machine learning, digital or computational humanities, or a related field with a substantial computational component

• Hands-on experience training, fine-tuning and analysing transformer language models in Python (PyTorch, Hugging Face), including on GPU clusters

• Experience with interpretability methods (feature attribution, probing, activation patching, sparse autoencoders) and/or with building evaluation datasets and benchmarks

• Experience designing experiments with proper controls, baselines and statistical care, and reproducible research-software practice

• A serious interest in historical texts, literary theory and the philosophy of language, and in reading model outputs closely alongside literary historians

• Strong English; knowledge of German is welcome

• Desirable: experience with historical or OCR-derived text or semantic change detection; annotation studies; released datasets or open-source tools; publications in NLP or ML venues; engagement with literary theory, historical materialism, hermeneutics, or structuralist or historical linguistics.