DylanJHJ/modernbert-base.cover-5k
0.1B • Updated • 104
Released checkpoints from "Search for Coverage: Learning Coverage-Aware Retrieval with Augmented Sub-Question Answerability"
Note This is the variant without relevance-based pre-fine-tuning. The results on BEIR is suboptimal (0.4835)
Note This checkpoint is for debugging, training for 25k on msmarco resulted in degraded performance on BEIR (0.4979)
Note Without any human-annotated labels, this checkpoint is built from scratch with SCOPE-flatten datasets. The training didn't help much, and the performance on BEIR remains similar (0.4763).