This report, posted on arXiv, investigates passage retrieval and re-ranking on the MS MARCO benchmark using the lightweight Qwen3-0.6B model. The authors examine performance across three experimental settings, including a BM25-then-reranking task and a corpus-wide dense retrieval task, with a third setting not fully described in the abstract.

The study focuses on two techniques: data curation and model souping. Data curation likely involves selecting or cleaning training data, while model souping typically refers to averaging weights of multiple fine-tuned models. The abstract does not specify the exact methods or results, but the combination is presented as a way to unlock the potential of the small model.

Given the truncated abstract, the full findings are not available. The report's contribution appears to be in demonstrating how these techniques can be applied to a compact model for a standard ranking benchmark.