OpenBMB releases Trident repository for mixed-modality retrieval
OpenBMB (ModelBest) published the Trident repository on GitHub, introducing a method to mitigate modality preference in mixed-modality retrieval. The repository is under construction; code, data, and model checkpoints are not yet available.
China context
- Original name
- OpenBMB
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
OpenBMB (ModelBest) has released the Trident repository on GitHub. Trident addresses modality preference in mixed-modality retrieval by building text, image, and fused text-image views of each document as co-equal positives and jointly optimizing relevance discrimination and positive-view balance via Multi-Positive View InfoNCE. The repository is under construction, with code, data, and model checkpoints to be released later.
Trident proposes a Multi-Positive View InfoNCE objective that treats text, image, and fused text-image views as co-equal positives. The method is claimed to improve mixed-modality retrieval on both CLIP-based and VLM-based architectures and reduce sensitivity to modality composition and text distractors. No code or checkpoints are available yet for verification.
Developers working on retrieval systems over mixed text-image corpora may face a new open-source option from OpenBMB once code and checkpoints are released, potentially reducing the need to build custom debiasing solutions. The repository currently offers no usable artifacts, so the immediate impact is limited to awareness of the approach.
For organizations building retrieval systems over mixed-modality document stores, Trident could eventually lower development costs by providing a pre-built method to handle modality bias. However, no license or release date is specified, so commercial use cannot be planned yet.
The next verifiable signal is the release of training code, evaluation code, and model checkpoints in the Trident repository. The arXiv paper is also listed as coming soon. Until these artifacts are published, the claimed improvements cannot be independently assessed.