For each dataset (set_covering, combinatorial auctions, CVS, IIS), we provide:
1. The codes for learning DIG-MILP to generate new instances.
To run the code, simply follow the numbers in the code name from `1' to `10' to finish training DIG-MILP.

2. The codes for downstream task - data sharing for solver tuning
Enter `downstream2' folder, and run the code inside the folder.
We also enclose the time_logs of the solver SCIP on DIG-MILP-generated instances and the original instances, through which we get the Pearson correlation coefficient between DIG-MILP-generated data and the original data in the paper.

3. The codes for downstream task - ML model training.
Enter `downstream' folder, and follow the order of the code number to train NNs to predict the optimal objective for MILPs. 
We also provide the trained DIG-MILP model dict trained on different $\gamma$ from which we get the results in the paper.

Data pre-processing for CVS and IIS could be more tedious, and thus we provide the primal format of the CVS data and the IIS data preprocessed into the primal format defined in the paper.

We publicly release all the codes of DIG-MILP.