Dataset proposed in the paper Building Dynamic Knowledge Graphs from Text-based Games.
This project is part of the project of Learning Dynamic Knowledge Graphs to Generalize on Text-based Games. Which we have released code in another repo.
| #Train | #Valid | #Test | Avg. Obs. | Avg. #Operations | #Vertices | #Edges | Avg. #Connections |
|---|---|---|---|---|---|---|---|
| 267,031 | 13,442 | 41,865 | 29.3 tokens | 3.1 | 99 | 10 | 43.1 |
We have split the dataset (dataset.zip, download) into training, validation, and test sets, each of them is in a .json file.
In a .json file, each line is a json dump, which is a single data point.
previous_triplets: ground truth (partially observable) graph at previous game step represented by triplets;previous_action: text action an agent performed at previous game step;observation: observation at current game step, which is resulted byprevious_action;target_commands: graph updates represented by text commands (operations).