brendenlake/MLC: Meta-Learning for Compositionality

This is an archived version of this github repository . Please see the repo for a README and to check for any updates.

Human behavioral data: The complete set of human and machine responses is available for easy viewing here (as HTML). Human behavioral data is available for download in the repository's 'data_human' folder. For few-shot instruction learning, the 'few_shot/val_behavior' subfolder has the 10 test trials with entries for different participants. For the open-ended task, the 29 participants are in 'open_ended_all/train' subfolder.

Model code: Meta-Learning for Compositionality (MLC) is an optimization procedure that encourages systematicity through a series of few-shot compositional tasks. This code shows how to train and evaluate a sequence-to-sequence (seq2seq) transformer in PyTorch to implement MLC for modeling human behavior.