vantuan5644/MLLM-CL-DCL

MLLM-CL domain continual learning data

This repo holds the domain continual learning (DCL) half of the MLLM-CL benchmark, with images unpacked and the DCL replay and router files next to it. It is a mirror of MLLM-CL/MLLM-CL and MLLM-CL/MLLM-CL-ReplayData, both Apache 2.0.

The upstream repo ships DCL together with ability continual learning (ACL), about 125 GiB in total. DCL alone is about 18 GiB. For ACL, use the upstream repo.

DCL has five domains, trained in this order: remote sensing (RS), medical (Med), autonomous driving (AD), science (Sci), finance (Fin).

Contents

path	what it is
MLLM-CL/domain/<domain>/{train,test}.json	instruction JSONs
MLLM-CL/domain/<domain>/images/	images, stored as tar shards
MLLM-CL-ReplayData/replay_DCL/task{1..5}.json	replay samples
MLLM-CL-ReplayData/router_DCL/task{1..5}.json	router training data

The layout matches upstream after its image tars are extracted. The image paths in domain/<domain>/*.json are relative to domain/<domain>/, for example images/10272.png. The paths in replay_DCL are relative to domain/, for example Med/images/image_15610.jpg.

Images are stored as uncompressed tar shards of up to 4 GiB (shard-NNNNN.tar). Tar member paths are relative to the repo root. MANIFEST.tsv.gz lists every file with its size and the shard that holds it.

Restore

Restore into the parent directory. You get MLLM-CL/ and MLLM-CL-ReplayData/ side by side.

pip install -U huggingface_hub
hf download --repo-type dataset vantuan5644/MLLM-CL-DCL restore.sh --local-dir .
bash restore.sh vantuan5644/MLLM-CL-DCL /path/to/datasets

restore.sh downloads the repo, extracts every shard in place, deletes the tars (set KEEP_TARS=1 to keep them), and checks each file in MANIFEST.tsv.gz for presence and size.

Citation

@article{zhao2025mllm,
  title={MLLM-CL: Continual Learning for Multimodal Large Language Models},
  author={Zhao, Hongbo and Zhu, Fei and Guo, Haiyang and Wang, Meng and Wang, Rundong and Meng, Gaofeng and Zhang, Zhaoxiang},
  journal={arXiv preprint arXiv:2506.05453},
  year={2025}
}