Octavius: Mitigating Task Interference in MLLMs via LoRA-MoE

Authors: Zeren Chen, Ziqin Wang, Zhen Wang, Huayang Liu, Zhenfei Yin, Si Liu, Lu Sheng, Wanli Ouyang, Jing Shao

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results (about 20% improvement) have shown the effectiveness and versatility of our design in various 2D and 3D downstream tasks.
Researcher Affiliation Collaboration 1 Shanghai AI Laboratory, 2 School of Software, Beihang University, 3 Institute of Artificial Intelligence, Beihang University, 4 University of Sydney
Pseudocode No The paper describes methods in prose and diagrams but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Code and datasets are available at https://openlamm.github.io/paper list/Octavius.
Open Datasets Yes Code and datasets are available at https://openlamm.github.io/paper list/Octavius.
Dataset Splits No The paper uses several well-known datasets but does not explicitly provide the training/validation/test splits for all of them needed for reproduction. For instance, it mentions 'fine-tune Octavius' and 'zero-shot evaluation' on various datasets but does not detail the splits used for training or validation across all datasets.
Hardware Specification Yes All experiments are conducted on 4 NVIDIA A100 80GB GPUs.
Software Dependencies No The paper mentions software components like Vicuna-13B, SentencePiece, and Adam optimizer, but does not provide specific version numbers for the underlying software libraries (e.g., PyTorch, TensorFlow, or SentencePiece library version) needed for reproducible setup.
Experiment Setup Yes The number of experts in the above three setups is 4, 3, and 6, respectively. The rank of each Lo RA expert is set to 32. During fine-tuning, we use an Adam (Kingma & Ba, 2014) optimizer with a total batch size of 64, a learning rate of 5 10 4, and an epoch of 4 on all setups.