CxrReportGen-Premium
MicrosoftAbout this model
The CXRReportGen Premium model is built on Qwen2.5-VL-7B-Instruct and is designed to help applications interpret complex chest X-ray imaging studies and generate radiology-style findings and impression. The model was trained using a two-stage process: first, supervised fine-tuning to adapt the foundation model to chest X-ray report generation, followed by reinforcement fine-tuning to further align outputs with preferred clinical report quality and structure. CXRReportGen Premium also supports parameter-efficient LoRA fine-tuning, enabling customers to further adapt the model to their own datasets, reporting conventions, and target deployment scenarios while preserving the base model architecture. When built upon and integrated into an application, CXRReportGen Premium may help developers generate comprehensive chest X-ray reports.
This repository contains the CXRReportGen Premium model, served via a vLLM Docker container with an inference middleware. Fine-tuned Lora adapters can be loaded at serving time without rebuilding the container.
This model is intended and provided as-is for research and model development exploration. CXRReportGen Premium is not designed or intended to be deployed in clinical settings as-is nor is it intended for use in the diagnosis or treatment of any health or medical condition (including generating radiology reports for use in patient care), and the model’s performance for such purposes has not been established. You bear sole responsibility and liability for any use of CXRReportGen Premium, including verification of outputs and incorporation into any product or service intended for a medical purpose or to inform clinical decision-making, compliance with applicable healthcare laws and regulations, and obtaining any necessary clearances or approvals.
Key model capabilities
CXRReportGen Premium supports LoRA fine-tuning via a BYOC (Bring Your Own Container) Docker image. The fine-tuning container uses the LlamaFactory framework and supports:
- LoRA rank up to 64
- Vision tower adaptation (LoRA applied to vision encoder)
- DeepSpeed ZeRO-2 for multi-GPU training
- OpenAI-format JSONL training data
Fine-tuned LoRA adapters can be loaded into the inference container at serving time via the LORA_PATH environment variable, without rebuilding the Docker image.
See Responsible AI for additional considerations for responsible use.
Key use cases
This model is intended and provided as-is for research and model development exploration
Out of scope use cases
CXRReportGen Premium is not designed or intended to be deployed in clinical settings as-is nor is it intended for use in the diagnosis or treatment of any health or medical condition (including generating radiology reports for use in patient care), and the model's performance for such purposes has not been established. You bear sole responsibility and liability for any use of CXRReportGen Premium, including verification of outputs and incorporation into any product or service intended for a medical purpose or to inform clinical decision-making, compliance with applicable healthcare laws and regulations, and obtaining any necessary clearances or approvals.