model-trajectories

This contains the model trajectories from What Do Language Models Learn and When? The Implicit Curriculum Hypothesis.

Huggingface Paper Page: https://huggingface.co/papers/2604.08510

Paper Link: https://arxiv.org/pdf/2604.08510

Citations

@article{liu2026language,
  title={What do Language Models Learn and When? The Implicit Curriculum Hypothesis},
  author={Liu, Emmy and Sun, Kaiser and Li, Millicent and Lee, Isabelle and Tjuatja, Lindia and Huang, Jen-tse and Neubig, Graham},
  journal={arXiv preprint arXiv:2604.08510},
  year={2026}
}

dataset_info: features:

• name: model dtype: string
• name: checkpoint dtype: string
• name: tokens_B dtype: int64
• name: task dtype: string
• name: task_type dtype: string
• name: metric dtype: string
• name: value dtype: float64 splits:
• name: train num_bytes: 782555 num_examples: 6985 download_size: 44115 dataset_size: 782555 configs:
• config_name: default data_files:
  • split: train path: data/train-* license: mit language:
• en