mousamoradi/qwen3-8b-glaucoma-reasoning-fold4

CISSGAN+Reasoner

CISSGAN+Reasoner is a public research demo for glaucoma clinical note analysis. It combines a CiSSGAN-based glaucoma classifier with a fine-tuned Qwen3-8B fold-4 reasoning adapter to predict glaucoma subtype and generate a concise one-sentence clinical justification.

Model Details

Model Description

This project is designed for glaucoma note classification and reasoning from free-text ophthalmology clinical notes. The pipeline first predicts a glaucoma subtype using the CiSSGAN classifier, then uses a fine-tuned Qwen3-8B LoRA adapter to generate a short clinical reasoning sentence aligned with the predicted subtype.

• —Developed by: Mousa Moradi
• —Created on: 4/21/2026 at 15:14:20
• —Contact: mmoradi2@meei.harvard.edu
• —Model type: Glaucoma clinical note classification and reasoning pipeline
• —Language(s) (NLP): English
• —License: MIT
• —Finetuned from model: Qwen/Qwen3-8B

Model Sources

• —Base model: Qwen/Qwen3-8B
• —Library: PEFT
• —Demo: Hugging Face Space: CISSGAN+Reasoner

Uses

Direct Use

This model is intended for research use on de-identified ophthalmology clinical notes. It can be used to:

• —classify glaucoma subtype from note text
• —provide a binary glaucoma versus non-glaucoma assessment
• —generate a one-sentence clinical reasoning statement for the predicted subtype
• —assist exploratory research on automated ophthalmic note interpretation

Downstream Use

Potential downstream uses include:

• —research prototypes for glaucoma note triage
• —retrospective analysis of de-identified note datasets
• —educational demonstrations of clinical NLP and reasoning systems
• —evaluation of classification-reasoning pipelines for ophthalmology AI

Out-of-Scope Use

This model is not intended for:

• —real-time clinical decision-making without physician oversight
• —diagnosis or treatment recommendation in routine patient care
• —use on identifiable protected health information in public deployments
• —emergency screening or unsupervised medical triage
• —use outside glaucoma-related ophthalmology note interpretation without further validation

Bias, Risks, and Limitations

This model is trained and evaluated in a specific research context and may not generalize to all institutions, note styles, patient populations, or documentation templates. Predictions and reasoning may reflect biases in the source data, labeling conventions, and subtype prevalence.

The reasoning output is generated conditioned on the predicted subtype, so it should be interpreted as a justification of the model prediction rather than an independent second opinion. Incorrect subtype predictions may therefore lead to plausible but incorrect reasoning text.

This system should be treated as a research tool only.

Recommendations

Users should:

• —use only de-identified or synthetic clinical notes in the public demo
• —manually review all outputs
• —avoid using outputs as a substitute for expert ophthalmic judgment
• —validate performance locally before any downstream research deployment
• —document institutional and population differences when evaluating generalization

How to Get Started with the Model

This project uses a fine-tuned PEFT adapter on top of Qwen/Qwen3-8B.

Example loading pattern:

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_name = "Qwen/Qwen3-8B"
adapter_path = "YOUR_ADAPTER_REPO_OR_LOCAL_PATH"

tokenizer = AutoTokenizer.from_pretrained(adapter_path)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype="auto",
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, adapter_path)
model.eval()