Adapter Based Domain Adaptation For Vision Language Tasks

via “fine-tuning and domain adaptation for specialized similarity tasks”

sentence-similarity model by undefined. 19,27,050 downloads.

Unique: Supports fine-tuning on the Qwen3-VL-2B-Instruct architecture with flexible loss functions and parameter-efficient approaches (LoRA, adapters), enabling domain adaptation without full model retraining while maintaining the unified multimodal embedding space

vs others: More efficient than training multimodal models from scratch because it leverages pre-trained vision and language components, reducing fine-tuning time by 10-50x and requiring significantly less labeled data (100s vs 100Ks of pairs)