All data is available in the data/ directory, whose subdirectories correspond to the three domains in SCONE (Alchemy, Scene, and Tangrams). Each subdirectory contains JSON files storing the processed train, dev, and test data. We also provide a script that generates action sequences for supervised learning (action_sequences.py).

Structure of the JSON files:

• One line per file, containing a single JSON object with a list of examples.
• Each example is a dictionary with a few properties:
  • identifier: A unique identifier for the interaction.
  • initial_env: The initial world state for the interaction.
  • utterances: Sequences of instructions. The instruction is the natural language instruction; after_env is the gold environment after executing the instruction according to the original data; and actions is a sequence of low-level actions generated from action_sequences.py and only used during the supervised learning setting.

See model/. There are a few necessary steps to start running the code.

• DyNet (see install guide for Python)
• Crayon (see install guide for use with DyNet)
• Docker

If you don't want to use Crayon for logging, you can disable it in the code.