GlucoMind-8B-Think

GlucoMind-8B-Think is an 8B-parameter Qwen3-based language model fine-tuned for personalized continuous glucose monitoring (CGM) recommendation report generation. It is designed to reason over structured patient profiles derived from CGM trajectories, meal and diet records, insulin or medication information, and clinical notes, then generate prioritized glucose-management recommendations.

The model is associated with the manuscript draft "GlucoMind-Think: Benchmarking and advancing large language models for personalized continuous glucose monitoring recommendation report generation". The draft introduces GlucoBench, a clinician-annotated benchmark for evaluating CGM recommendation report generation through insight identification, etiology analysis, and intervention recommendation.

This model is intended for research and clinician-in-the-loop decision-support prototyping. It is not a medical device and must not be used as a substitute for professional medical judgment, diagnosis, treatment, medication adjustment, or emergency care.

Model Details

• Model name: GlucoMind-8B-Think
• Base model: Qwen/Qwen3-8B
• Architecture: Qwen3ForCausalLM
• Model type: decoder-only causal language model
• Fine-tuning method: supervised fine-tuning with structured chain-of-thought supervision
• Training focus: CGM-based clinical concept reasoning and recommendation prioritization
• Precision: bfloat16
• Context length: 40,960 tokens
• Checkpoint format: safetensors, sharded into 4 weight files
• Recommended Transformers version: transformers>=4.51

Intended Use

This model is intended for research on CGM-based recommendation report generation, candidate recommendation ranking, structured CGM interpretation drafts, and medical-reasoning experiments over glucose trajectories and patient context.

The expected input is a structured textual patient profile, not raw CGM sensor files. The profile should summarize relevant glucose trajectories, meal timing and composition, insulin or medication context, and other clinical notes available for the case.

Out-of-Scope Use

Do not use this model for autonomous diabetes management, medical diagnosis, treatment decisions, insulin or medication dose adjustment without qualified clinician oversight, emergency triage, or regulated clinical use without separate validation and approval.

Training Data

GlucoMind-8B-Think was fine-tuned on structured chain-of-thought samples constructed from the training split of GlucoBench. Patient cases are represented as structured textual profiles derived from CGM time series, diet records, insulin or medication information, and clinical notes.

For supervised fine-tuning, a teacher LLM was used to generate structured reasoning traces linking patient-specific evidence to clinical concepts and recommendation decisions. After quality control, 3,757 high-quality chain-of-thought samples were retained for fine-tuning.

The training data and benchmark annotations are not included in this model repository.

How to Use

For the full codebase, preprocessing pipeline, prompts, inference examples, and evaluation scripts, please refer to the GitHub repository:

https://github.com/ADreamPen/GlucoMind-Think

Basic loading example:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ADreamPen/GlucoMind-8B-Think"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

The model uses the Qwen3 chat template. For CGM recommendation report generation, provide a structured patient profile containing glucose trajectory summaries, meal context, insulin or medication information, and relevant clinical notes.

Ethical and Safety Considerations

CGM recommendation reports can influence health decisions. The model output should be reviewed by qualified clinicians and used only as a research or decision-support draft. Users are responsible for validating factuality, safety, and clinical appropriateness before using the model in any health-related workflow.

License

This model is released under the Apache 2.0 license, following the license of the Qwen3-8B base model. Users are responsible for ensuring that their use complies with the license terms and all applicable laws, regulations, data governance requirements, and clinical safety requirements.