xingshen/prompt4trust-cgpgenerator-1.5B
Prompt4Trust-CGPGenerator-1.5B by xingshen is a 1.5 billion parameter, 32K context length language model based on Qwen2.5-1.5B-Instruct. It is specifically fine-tuned as a Calibration Guidance Prompt Generator within the Prompt4Trust framework. This model generates context-aware auxiliary prompts to improve confidence calibration and trustworthiness in multimodal large language models (MLLMs) for healthcare applications, achieving state-of-the-art results on the PMC-VQA benchmark.
Loading preview...
Prompt4Trust-CGPGenerator-1.5B: Calibrating MLLM Confidence for Healthcare
xingshen/prompt4trust-cgpgenerator-1.5B is a specialized 1.5 billion parameter language model, fine-tuned from Qwen2.5-1.5B-Instruct, designed to enhance the trustworthiness of Multimodal Large Language Models (MLLMs) in safety-critical healthcare settings. Developed as part of the Prompt4Trust framework, this model acts as a Calibration Guidance Prompt Generator.
Key Capabilities
- Confidence Calibration: Generates context-aware auxiliary prompts to guide downstream MLLMs, ensuring their confidence scores more accurately reflect true prediction accuracy.
- Reinforcement Learning Framework: Utilizes a reinforcement learning approach for prompt augmentation, focusing on clinically meaningful calibration.
- Improved Reliability: Addresses challenges of prompt sensitivity and overconfident incorrect responses in MLLMs.
- Enhanced Task Performance: Improves both the reliability and overall task performance of MLLMs, achieving state-of-the-art results on the PMC-VQA benchmark.
- Efficient Generalization: Enables efficient zero-shot generalization to larger MLLMs.
Good for
- Developers and researchers working on trustworthy AI in healthcare.
- Applications requiring calibrated confidence scores from MLLMs.
- Improving the reliability and safety of MLLM deployments in clinical environments.
- Augmenting MLLM prompts to reduce overconfidence and enhance accuracy.