We present ‘gradiend‘, an open-source Python package that operationalizes the GRADIEND method for learning feature directions from factual-counterfactual MLM and CLM gradients in language models. The package provides a unified workflow for feature-related data creation, training, evaluation, visualization, persistent model rewriting via controlled weight updates, and multi-feature comparison. We demonstrate ‘gradiend‘ through an English pronoun running example, a semantic sentiment use case that evaluates lexical generalization to held-out target words, and a large-scale feature comparison.

Language models perform well on grammatical agreement, but it is unclear whether this reflects rule-based generalization or memorization. We study this question for German definite singular articles, whose forms depend on gender and case. Using GRADIEND, a gradient-based interpretability method, we learn parameter update directions for gender-case specific article transitions. We find that updates learned for a specific gender-case article transition frequently affect unrelated gender-case settings, with substantial overlap among the most affected neurons across settings. These results argue against a strictly rule-based encoding of German definite articles, indicating that models at least partly rely on memorized associations rather than abstract grammatical rules.

This paper describes an approach to automat- ically close the knowledge gap of Clickbait- Posts via a transformer model trained for Question-Answering, augmented by a task- specific post-processing step. This was part of the SemEval 2023 Clickbait shared task (Frbe et al., 2023a) - specifically task 2. We devised strategies to improve the existing model to fit the task better, e.g. with different special mod- els and a post-processor tailored to different inherent challenges of the task. Furthermore, we explored the possibility of expanding the original training data by using strategies from Heuristic Labeling and Semi-Supervised Learn- ing. With those adjustments, we were able to improve the baseline by 9.8 percentage points to a BLEU-4 score of 48.0%.