According to the research paper, the following line : "We directly adopted the evaluator again for knowledge strips filtering, and the top-k is set to 5, filter threshold as -0.5", mentions that a threshold of -0.5 was used to filter the strips, but it is nowhere to be found in the internal_knowledge_preparation.py script.