metadata
language: en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- multimodal
- audio
- healthcare
- respiratory
- question-answering
- mixture-of-experts
- lora
RAMoEA-QA (Checkpoint)
RAMoEA-QA is a hierarchical generative model for Respiratory Audio Question Answering (RA-QA). It supports multiple question formats (open-ended, single-verify, multiple-choice) and both discrete labels (e.g., diagnosis/verification) and continuous targets (regression) within a single system.
Architecture (two-stage conditional specialization):
- Audio Mixture-of-Experts (Audio-MoE): routes each (audio, question) example to one pre-trained audio encoder expert.
- Language Mixture-of-Adapters (MoA): selects one LoRA adapter on a shared frozen LLM backbone (GPT-2) to match query intent and answer format.
Intended use
Research and decision-support experiments on respiratory-audio question-answering. Not a medical device.
Usage
This checkpoint is meant to be used with the accompanying codebase (audio encoder factory + routing + alignment):