How does splitting a query into k per-chunk training ex...

When you split a query into k chunks, the training distribution is heavily skewed toward insufficient context. If the recall is r, the fraction of supporting chunks is r/k. In a scenario where r=0.8 and k=10, the supporting fraction is only 8%. This creates an aggressive imbalance that must be corrected by either weighting the supporting instances (e.g., a weight of 17.25) or subsampling the non-supporting instances to ensure the model does not over-index on negative examples.