HELMET

An application-centric long-context evaluation suite from Danqi Chen's princeton-nlp group (ICLR 2025, with Intel): HELMET moves beyond synthetic needle-in-a- haystack tests to diverse, realistic long-context tasks with reliable, model-comparable scoring. It has been adopted in major model reports including Llama-4 and Jamba. Companion to the group's ProLong long-context training work.