README.md

Box Dataset

These directiories contain all the data splits used in the paper Entity Tracking in Language Models.

Splits:

• T5 datasets (used for fine-tuning and evaluating T5):

  • t5_boxes_nso_exp2_max3: Base and Vocab splits.
  • t5_boxes_nso_numops_trainlen=2_exp2_max3: NumOps split.
  • t5_boxes_nso_exp2_max3_alt_forms_train: AltForms and AltForms+NumOps splits.
  • t5_boxes_nso_exp2_max3_move_contents: MoveContents split.
  • t5_boxes_nso_exp2_max3_ambiref: AmbiRef split.

• Zero-shot/few-shot datasets (used for in-context experiments)

  • few_shot_boxes_nso_exp2_max3: Base split.
  • few_shot_boxes_nso_exp2_max3_move_contents: MoveContents split.
  • few_shot_boxes_nso_exp2_max3_ambiref AmbiRef split.

See the README.md file in each directory for a more detailed description of each split.