Explain the non-monotonic effect of post-training corpus size

Determine whether the non-monotonic relationship between post-training corpus size and ComCLIP performance is caused by single-epoch training dynamics, data quality, or a model-capacity limit.

Background

The experiments compare ComCLIP post-training on CC3M, CC12M, and COCO Caption. Increasing the corpus from CC3M to CC12M does not consistently improve performance: for example, zero-shot accuracy is similar while linear-probe performance can decline.

The paper reports this behavior but does not identify its cause. The authors explicitly formulate an unresolved question concerning whether the effect arises from the single-epoch schedule, differences in data quality, or the capacity of the model being refined.

References

Performance is not monotone in corpus size (Sec.~\ref{sec:exp-main}); whether this reflects single-epoch dynamics, data quality, or a capacity limit is open.

— Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder  (2610.11374 - Wang et al., 8 Oct 2026) in Section 5, “Limitations and Future Work,” item (iv)