Entity Tracking in Language Models
Paper
•
2305.02363
•
Published
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.