lablab-ai-amd-developer-hackathon/asd-interpreter-merged
The lablab-ai-amd-developer-hackathon/asd-interpreter-merged is a 7.6 billion parameter clinical language interpreter fine-tuned from Qwen/Qwen2.5-7B-Instruct. Developed by lablab.ai and Raghav Aryen on AMD MI300X hardware, it specializes in generating patient-facing clinical summaries from fMRI connectivity reports for Autism Spectrum Disorder (ASD). This model processes structured prompts containing ASD probabilities and network-level gradient saliency scores to output detailed clinical interpretations.
Loading preview...
ASD Clinical Interpreter
This model, developed by lablab.ai and Raghav Aryen, is a specialized clinical language interpreter fine-tuned from Qwen/Qwen2.5-7B-Instruct with 7.6 billion parameters. It was trained using QLoRA on AMD MI300X hardware (ROCm 7.0) and is designed to translate complex fMRI connectivity reports into understandable clinical summaries for Autism Spectrum Disorder (ASD).
Key Capabilities
- Clinical Summary Generation: Produces concise, patient-facing clinical summaries from structured inputs.
- Detailed Interpretation: Outputs include overall impression, confidence levels, identification of influential brain networks, site-invariance assessment, and recommended next steps for clinical review.
- Specialized Input Processing: Takes ensemble ASD probability, per-model predictions from 20 LOSO GCN models, and network-level gradient saliency scores (e.g., DMN, Salience, Frontoparietal).
- Integration: Used live in the BrainConnect-ASD Space to generate reports from GCN ensemble inference.
Training and Limitations
- Training Data: Fine-tuned on synthetic clinical summaries derived from ABIDE I gradient saliency outputs, manually curated for clinical tone.
- Hardware: Trained on AMD MI300X (192GB HBM3) with ROCm 7.0.
- Context Length: Supports a context length of 4096 tokens.
- Limitations: Not validated on real clinical populations and not intended as a medical device. Performance may degrade on atlases other than CC200.