[CuriousJ]

This paper shows that 200-800 is the ideal chunk size; if you go above, the model starts getting confused / distracted. https://arxiv.org/pdf/2406.14497

[zaptrem]

Makes sense. Thanks!