This repository is a forked and modified version of the original TopiCK (EMNLP 2025) implementation. The goal of this fork is to provide a robust and reproducible experimental setup by addressing practical issues in inference, evaluation, and model configuration.
This fork introduces the following changes on top of the original implementation:
-
Stabilized inference pipeline
- Revised prediction handling to correctly support label-based outputs
- Fixed evaluation logic to ensure consistent and reliable accuracy computation
-
Controlled model configuration
- Adapted the codebase to support lightweight model variants
- Simplified experimental settings for reproducible local runs
-
Improved retrieval stability
- Fixed FAISS import and runtime issues observed during retrieval
- Reduced environment-specific failure cases
These changes focus on improving experimental reliability and reproducibility rather than maximizing peak model performance.
This repository is forked from the original implementation of TopiCK (EMNLP 2025). All credit for the original method and experimental design belongs to the original authors.
This fork contains only practical modifications for stable experimentation and reproducible evaluation.