README.md

metadata

configs:
  - config_name: mcd1
    data_files:
      - split: train
        path: mcd1/train-*
      - split: dev
        path: mcd1/dev-*
      - split: test
        path: mcd1/test-*
  - config_name: mcd2
    data_files:
      - split: train
        path: mcd2/train-*
      - split: dev
        path: mcd2/dev-*
      - split: test
        path: mcd2/test-*
  - config_name: mcd3
    data_files:
      - split: train
        path: mcd3/train-*
      - split: dev
        path: mcd3/dev-*
      - split: test
        path: mcd3/test-*
dataset_info:
  - config_name: mcd1
    features:
      - name: commands
        dtype: string
      - name: actions
        dtype: string
    splits:
      - name: train
        num_bytes: 1435200
        num_examples: 8365
      - name: dev
        num_bytes: 242915
        num_examples: 1046
      - name: test
        num_bytes: 249212
        num_examples: 1045
    download_size: 340627
    dataset_size: 1927327
  - config_name: mcd2
    features:
      - name: commands
        dtype: string
      - name: actions
        dtype: string
    splits:
      - name: train
        num_bytes: 1408018
        num_examples: 8365
      - name: dev
        num_bytes: 229805
        num_examples: 1046
      - name: test
        num_bytes: 230998
        num_examples: 1045
    download_size: 336499
    dataset_size: 1868821
  - config_name: mcd3
    features:
      - name: commands
        dtype: string
      - name: actions
        dtype: string
    splits:
      - name: train
        num_bytes: 1419109
        num_examples: 8365
      - name: dev
        num_bytes: 252766
        num_examples: 1046
      - name: test
        num_bytes: 247900
        num_examples: 1045
    download_size: 340622
    dataset_size: 1919775

Dataset Card for "SCAN_MCDSplits"

This is the dataset repository for SCAN MCD splits. In total, there are three splits - mcd1, mcd2, and mcd3

SCAN is a set of simple language-driven navigation tasks for studying compositional learning and zero-shot generalization. The SCAN tasks were inspired by the CommAI environment, which is the origin of the acronym (Simplified versions of the CommAI Navigation tasks).

The relevant SCAN paper is:

Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. ICML 2018.

The relevant MCD split paper is:

Measuring Compositional Generalization: A Comprehensive Method on Realistic Data. ICLR 2020.

You can load them by: datasets.load_dataset("Punchwe/SCAN_MCDSplits", name="mcd1", split="train")