← All questions
Retrieval & EmbeddingsHard· SPLADE· Performance· Sparsity

Why might two SPLADE checkpoints with the same average...

The cost is proportional to the sum of query activations multiplied by document activations. The two models can differ by up to the size of the vocabulary in cost even if the average count is the same. To diagnose, dump the per-dimension activation frequencies and compare the effective measure (meff). The slow model often results from an L1 penalty or a FLOPS penalty applied incorrectly, which removes the necessary cross-document coupling.

Code playground (Python)