AdaptLLM/biomed-Qwen2.5-VL-3B-Instruct
AdaptLLM/biomed-Qwen2.5-VL-3B-Instruct is a 3 billion parameter multimodal large language model (MLLM) developed by AdaptLLM, specifically adapted from Qwen2.5-VL-3B-Instruct for biomedicine. This model is post-trained on a specialized dataset to enhance its performance in biomedical visual instruction tasks. It is designed for applications requiring visual question answering and understanding within the biomedical domain, leveraging its base architecture's multimodal capabilities.
Loading preview...
AdaptLLM/biomed-Qwen2.5-VL-3B-Instruct: Domain-Adapted Multimodal LLM
This model is a 3 billion parameter multimodal large language model (MLLM) developed by AdaptLLM, specifically fine-tuned from the Qwen2.5-VL-3B-Instruct base model. Its primary distinction lies in its domain-adaptive post-training for biomedicine, as detailed in the EMNLP 2025 paper "On Domain-Adaptive Post-Training for Multimodal Large Language Models" (arXiv:2411.19930).
Key Capabilities & Features
- Biomedical Domain Specialization: Optimized for understanding and responding to queries related to biomedical images and text.
- Multimodal Input: Inherits the Qwen2.5-VL architecture, allowing for combined image and text inputs.
- Instruction Following: Designed to follow instructions for tasks like image description and visual question answering in a biomedical context.
- Reproducibility: The training dataset, biomed-visual-instructions, is publicly available, enabling reproduction of the model using LlamaFactory.
When to Use This Model
- Biomedical Visual QA: Ideal for applications requiring analysis and interpretation of biomedical images with accompanying textual queries.
- Domain-Specific MLLM Research: Useful for researchers exploring domain adaptation techniques for multimodal models.
- Healthcare AI Development: Suitable for integrating MLLM capabilities into healthcare-related tools that process visual data.
This model was trained for one epoch with a learning rate of 1e-5 and a global batch size of 128, building upon the robust Qwen2.5-VL-3B-Instruct foundation.