|
--- |
|
license: mit |
|
tags: |
|
- MAPF |
|
pretty_name: MAPF-GPT |
|
size_categories: |
|
- 100M<n<1B |
|
--- |
|
# Dataset Card for Dataset Name |
|
|
|
<!-- Provide a quick summary of the dataset. --> |
|
|
|
This dataset contains 1 billion observation-action pairs to clone the behavior of LaCAM approach to solve MAPF problem via transformer (GPT) model. |
|
|
|
## Dataset Details |
|
|
|
- **Curated by:** Anton Andreychuk ([@aandreychuk](https://huggingface.co/aandreychuk/)), Alexey Skrynnik ([@tviskaron](https://huggingface.co/tviskaron/)) |
|
- **License:** MIT |
|
- **Repository:** [GitHub](https://github.com/Cognitive-AI-Systems/MAPF-GPT) |
|
- **Model:** [MAPF-GPT](https://huggingface.co/aandreychuk/MAPF-GPT/tree/main) |
|
- **Paper:** [ArXiv](https://arxiv.org/abs/2409.00134) |
|
|
|
## Dataset Structure |
|
|
|
Dataset contains `train` and `validation` parts. |
|
- `train` part contains 1,000 * 2^20 obsevation-action pairs. They are divided into 500 `.arrow` files. |
|
- `validation` part contains 2^20 observation-action pairs. They are saved into a single `.arrow` file. |
|
Dataset requires 256 GB of disk space. |
|
|
|
More details about creation of the dataset, the source data, the structure of observation, etc. are provided in the [paper](https://arxiv.org/abs/2409.00134) |
