README.md

# DataBack: Dataset of SAT Formulas and Backbone Variable Phases

## What is DataBack
`DataBack` is a dataset that consists of SAT formulas in CNF format, each labeled with the phases of its backbone variables. Within `DataBack`, there are two distinct subsets: the pre-training set, named `DataBack-PT`, and the fine-tuning set, named `DataBack-FT`. The state-of-the-art backbone extractor called [`CadiBack`](https://wenxiwang.github.io/papers/cadiback.pdf) has been employed to obtain the backbone variable phases. We have allocated a timeout of 1,000 seconds for the pre-training formulas and 5,000 seconds for the fine-tuning ones due to the increased complexity of the fine-tuning formulas.

The `DataBack` dataset has been employed to both pre-train and fine-tune our `NeuroBack` model, which has demonstrated significant improvements in SAT solving efficiency. For an in-depth exploration of `DataBack`, please refer to our [`NeuroBack`](https://arxiv.org/pdf/2110.14053.pdf) or [`CadiBack`](https://wenxiwang.github.io/papers/cadiback.pdf) papers.

## Contributors
Wenxi Wang ([email protected]), Yang Hu ([email protected])

## References
If you use `DataBack` in your research, please kindly cite the following papers.

[`NeuroBack`](https://arxiv.org/pdf/2110.14053.pdf) paper:
```bib
@article{wang2023neuroback,
  author = {Wang, Wenxi and
            Hu, Yang and
            Tiwari, Mohit and
            Khurshid, Sarfraz and
            McMillan, Kenneth L. and
            Miikkulainen, Risto},
  title = {NeuroBack: Improving CDCL SAT Solving using Graph Neural Networks},
  journal={arXiv preprint arXiv:2110.14053},
  year={2021}
}
```

[`CadiBack`](https://wenxiwang.github.io/papers/cadiback.pdf) paper:
```bib
@inproceedings{biere2023cadiback,
  title={CadiBack: Extracting Backbones with CaDiCaL},
  author={Biere, Armin and Froleyks, Nils and Wang, Wenxi},
  booktitle={26th International Conference on Theory and Applications of Satisfiability Testing (SAT 2023)},
  year={2023},
  organization={Schloss Dagstuhl-Leibniz-Zentrum f{\"u}r Informatik}
}
```

## Directory Structure
```
|- original             # Original SAT formulas and their backbone variable phases
|    |- cnf_pt.tar.gz   # SAT formulas (in CNF format) for model pre-training
|    |- bb_pt.tar.gz    # Backbone phases for pre-training formulas
|    |- cnf_ft.tar.gz   # SAT formulas (in CNF format) for model fine-tuning
|    |- bb_ft.tar.gz    # Backbone phases for fine-tuning formulas
|
|- dual                      # Dual SAT formulas and their backbone variable phases
|    |- dual_cnf_pt.tar.gz   # Dual SAT formulas (in CNF format) for model pre-training
|    |- dual_bb_pt.tar.gz    # Backbone phases for dual pre-training CNFs
|    |- dual_cnf_ft.tar.gz   # Dual SAT formulas (in CNF format) for model fine-tuning
|    |- dual_bb_ft.tar.gz    # Backbone phases for dual fine-tuning CNFs
```

## File Naming Convention
In the original directory, each CNF tar file (`cnf_*.tar.gz`) contains compressed CNF files named: `[cnf_name].[compression_format]`, where `[compression_format]` could be bz2, lzma, xz, gz, etc. Correspondingly, each backbone tar file (`bb_*.tar.gz`) comprises compressed backbone files named: `[cnf_name].backbone.xz`. It is important to note that a compressed CNF file will always share its `[cnf_name]` with its associated compressed backbone file.

In the dual directory, the naming convention remains consistent, but with an added `dual_` prefix for each CNF tar file or backbone tar file, and an `d_` prefix for each compressed CNF or backbone file to indicate that it pertains to a dual SAT formula.

## Format of the Extracted Backbone File 
The extracted backbone file (`*.backbone`) adheres to the output format of [`CadiBack`](https://wenxiwang.github.io/papers/cadiback.pdf). Specifically, each line contains the letter 'b' followed by an integer. This integer represents the backbone variable ID and its associated phase.