metadata
license: mit
task_categories:
- text-generation
dataset_info:
features:
- name: preds
sequence: string
- name: rules
sequence:
- name: premises
sequence: string
- name: conclusion
dtype: string
- name: facts
sequence: string
- name: query
dtype: string
- name: label
dtype: int64
- name: depth
dtype: int64
splits:
- name: train_lp_downsampled
num_bytes: 659503714
num_examples: 700000
- name: train_rp_balanced_downsampled
num_bytes: 850381210
num_examples: 700000
- name: validation_lp
num_bytes: 6625773
num_examples: 7000
- name: validation_lp_balanced
num_bytes: 7651259
num_examples: 7000
- name: validation_lp_balanced_deep_30_pred
num_bytes: 15515151
num_examples: 13000
- name: validation_lp_balanced_deep_60_pred
num_bytes: 27707171
num_examples: 13000
- name: validation_lp_star_balanced_deep_30_pred
num_bytes: 16249806
num_examples: 13000
- name: validation_lp_star_balanced_deep_60_pred
num_bytes: 25884528
num_examples: 13000
- name: validation_rp
num_bytes: 7501908
num_examples: 7000
- name: validation_rp_balanced
num_bytes: 8551772
num_examples: 7000
- name: validation_rp_balanced_deep_30_pred
num_bytes: 19016725
num_examples: 13000
- name: validation_rp_balanced_deep_60_pred
num_bytes: 32788646
num_examples: 13000
- name: validation_rp_balanced_3_premise
num_bytes: 11934368
num_examples: 7000
- name: validation_rp_balanced_2_premise
num_bytes: 8774321
num_examples: 7000
- name: validation_rp_balanced_1_premise
num_bytes: 5474611
num_examples: 7000
- name: validation_rp_balanced_1_2_premise
num_bytes: 6996474
num_examples: 7000
- name: validation_rp_balanced_2_3_premise
num_bytes: 10312203
num_examples: 7000
download_size: 268682683
dataset_size: 1720869640
configs:
- config_name: default
data_files:
- split: train_lp_downsampled
path: data/train_lp_downsampled-*
- split: train_rp_balanced_downsampled
path: data/train_rp_balanced_downsampled-*
- split: validation_lp
path: data/validation_lp-*
- split: validation_lp_balanced
path: data/validation_lp_balanced-*
- split: validation_lp_balanced_deep_30_pred
path: data/validation_lp_balanced_deep_30_pred-*
- split: validation_lp_balanced_deep_60_pred
path: data/validation_lp_balanced_deep_60_pred-*
- split: validation_lp_star_balanced_deep_30_pred
path: data/validation_lp_star_balanced_deep_30_pred-*
- split: validation_lp_star_balanced_deep_60_pred
path: data/validation_lp_star_balanced_deep_60_pred-*
- split: validation_rp
path: data/validation_rp-*
- split: validation_rp_balanced
path: data/validation_rp_balanced-*
- split: validation_rp_balanced_deep_30_pred
path: data/validation_rp_balanced_deep_30_pred-*
- split: validation_rp_balanced_deep_60_pred
path: data/validation_rp_balanced_deep_60_pred-*
- split: validation_rp_balanced_3_premise
path: data/validation_rp_balanced_3_premise-*
- split: validation_rp_balanced_2_premise
path: data/validation_rp_balanced_2_premise-*
- split: validation_rp_balanced_1_premise
path: data/validation_rp_balanced_1_premise-*
- split: validation_rp_balanced_1_2_premise
path: data/validation_rp_balanced_1_2_premise-*
- split: validation_rp_balanced_2_3_premise
path: data/validation_rp_balanced_2_3_premise-*
This dataset accompanies the paper The Scaling Properties of Implicit Deductive Reasoning in Transformers. It contains synthetic Horn clause reasoning problems with features such as rules, facts, queries, and labels, designed to study implicit deductive reasoning in depth-bounded Transformers.
Code for dataset generation, model training, and evaluation is available at: https://github.com/envomp/Implicit-Deductive-Reasoning-in-Transformers