Mannyboyy/

randomized-yarn-mrcr

Frozen, per-model binned train/test splits for the MRCR experiments in Randomized YaRN Improves Length Generalization for Long-Context Reasoning (Mehta, Yin, Durrett). MRCR samples are sorted into length bins by token count, which is tokenizer-dependent — so each model gets its own config. Published as fixed bytes so the paper reproduces independently of transformers versions.

	Configs / splits




	

config tokenizer train test

qwen… See the full description on the dataset page: https://huggingface.co/datasets/Mannyboyy/randomized-yarn-mrcr.