jesserem/cdqac_datasets

CDQAC offline datasets for job shop scheduling

Training datasets, evaluation instances and pretrained checkpoints for Generalizing Beyond Suboptimality: Offline Reinforcement Learning Learns Effective Scheduling through Random Solutions (TMLR 2026). Code: jesserem/CDQAC_scheduling.

Contents

folder	contents
train_datasets/<problem>/<name>/	offline training datasets (see below)
train_datasets/legacy/	the same training data in the original single-file .npy format
eval_datasets/	evaluation instances: SD1 and SD2 test sets (10x5 to 40x10), FJSP benchmarks (Brandimarte, Hurink edata/rdata/vdata), JSP benchmarks (Taillard, DMU, Lawrence)
checkpoints/	pretrained actors of CDQAC and the baselines (BC, IQL, TD3-AWR, Offline-LD)

Training datasets

id	problem	train / val instances	operations per instance
fjsp/sd1_10x5	FJSP, 10 jobs x 5 machines	1000 / 100	52 / 44-55
fjsp/sd1_15x10	FJSP, 15 x 10	1000 / 100	155 / 141-156
fjsp/sd1_20x10	FJSP, 20 x 10	500 / 100	199 / 184-216
jsp/taillard_10x5	JSP, 10 x 5	500 / 100	50
jsp/taillard_15x10	JSP, 15 x 10	500 / 100	150

Each dataset directory contains:

• dataset.json: what the dataset holds, how it was generated (seeds) and sha256 checksums
• train.npz, val.npz: the instances: job_length [N, n_j], op_pt [N, O_max, n_m] (processing time of each operation on each machine, 0 = not allowed; rows past n_ops are padding) and n_ops [N]
• sources/<source>.npz: trajectories for the training instances, as actions [N_s, K, O] (action = job * n_m + machine) and makespan [N_s, K]

source	trajectories per instance
pdr	16 priority dispatching rules named <job>_<machine> (job rule {MWR, LWR, LOPNR, MOPNR} x machine rule {EST, LST, LPT, SPT}, e.g. MWR_EST); 4 for JSP
random	300 uniformly random rollouts
ga_hof	100 genetic-algorithm hall-of-fame solutions, sorted by makespan
ga_pop	200 members of the final genetic-algorithm population (with duplicates)
eps_pdr_0.1, eps_pdr_0.2	100 dispatching-rule rollouts that take a random action with probability 0.1 / 0.2 (FJSP only; first 500 instances)

All trajectories satisfy the environment's action mask: at every decision they only pick an operation-machine pair that can start at the current decision time. Each dataset.json declares "action_mask": true, and the code always trains and evaluates with this mask.

All files are plain NumPy arrays; loading them needs no allow_pickle:

import numpy as np

d = "train_datasets/fjsp/sd1_10x5"
train = np.load(f"{d}/train.npz")            # job_length, op_pt, n_ops
random = np.load(f"{d}/sources/random.npz")  # actions, makespan

The code repository documents the format in cdqac/data/README.md and trains on a dataset with python train.py --dataset fjsp/sd1_10x5.

Download

from huggingface_hub import snapshot_download

snapshot_download("jesserem/cdqac_datasets", repo_type="dataset", local_dir=".",
                  allow_patterns=["train_datasets/fjsp/sd1_10x5/*"])

or, from the code repository, python download_datasets.py fjsp/sd1_10x5.

Legacy files

train_datasets/legacy/{fjsp,jsp}/ holds the original single-file datasets (train_10_5.npy, val_10_5.npy, ...): pickled arrays of per-instance dicts with the same data. Besides the 16 dispatching rules above (named there with a _masked suffix, e.g. MWR_EST_masked), they also hold LCT/MCT/FAM rule variants, which the code does not use. For code released before the dataset directories, the legacy files are temporarily also kept at their original paths train_datasets/{fjsp,jsp}/*.npy; there, train_datasets/fjsp/val_10_5 (3).npy is the 20x10 validation set (legacy/fjsp/val_20_10.npy).

Citation

@article{remmerden2026generalizing,
    title={Generalizing Beyond Suboptimality: Offline Reinforcement Learning Learns Effective Scheduling through Random Solutions},
    author={Jesse van Remmerden and Zaharah Bukhsh and Yingqian Zhang},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2026},
    url={https://openreview.net/forum?id=4vVhm4817Q}
}