asadullahdogarr/gemma-3-4b-it-neglected-tropical-diseases

Gemma-3-4B-IT: Global Health Unstructured NTD Triage

This model is a fine-tuned LoRA adapter designed to act as an edge-deployable epidemiological diagnostic engine. It is specifically engineered to process highly volatile, multi-source, and unformatted health data to predict Neglected Tropical Disease (NTD) outbreaks.

Model Details

Model Description

Standard foundational models often fail when confronted with the reality of rural healthcare IT infrastructure. This model was fine-tuned using Adaption Labs' AutoScientist to bridge that gap. It ingests deeply unstructured contexts—including unparsed R&D funding string dumps, embedded JSON environmental telemetry, and code-mixed multilingual physician triage notes (Urdu, Spanish, English)—and outputs a clean, standardized clinical diagnosis alongside a data-backed justification.

Designed under the "Tiny AutoScientist" framework, this under-10B parameter model is explicitly optimized to fit onto a single local device, enabling off-grid field clinics to leverage advanced AI without requiring an internet connection.

• Developed by: Asad Ullah Dogar (Adaption Ambassador)
• Model type: Causal Language Model with LoRA Adapter
• Language(s) (NLP): English, Spanish, Urdu (Multilingual Triage Interpretation)
• License: Apache 2.0
• Finetuned from model: togethercomputer/gemma-3-4b-it

Uses

Direct Use

The model is intended to be used as an intelligent parsing and triage assistant for epidemiological data. It accepts noisy string inputs (like combined telemetry, date records, and localized doctor notes) and reliably formats them into a structured output comprising a Diagnosis and a Justification.

Downstream Use

Integration into offline, edge-deployed clinical dashboards or mobile health (mHealth) applications for rural healthcare workers managing diseases like Malaria, Dengue, Trachoma, and Buruli ulcer.

Out-of-Scope Use

This model is designed for data normalization and preliminary triage. It is not a substitute for definitive medical diagnosis by a certified healthcare professional. It should not be used to autonomously prescribe medication or determine critical care pathways without human-in-the-loop validation.

Bias, Risks, and Limitations

While the model has been equipped with strict hallucination mitigation guardrails during training, it relies heavily on historical OWID (Our World in Data) funding strings and localized telemetry. Anomalies or gaps in real-world telemetry (e.g., sensor errors) may impact the justification accuracy. Furthermore, the model's performance is tied to the clinical profiles of the specific NTDs it was trained on and may not generalize to novel pathogens.

Recommendations

Users should deploy this model strictly as an assistive data-structuring tool. Healthcare workers must verify the extracted justifications against their own on-the-ground clinical observations.

How to Get Started with the Model

Use the code below to load the adapter with the base model for local inference:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "togethercomputer/gemma-3-4b-it"
adapter_model_name = "your-username/gemma-3-4b-it-neglected-tropical-diseases" # Update with your exact repo path

tokenizer = AutoTokenizer.from_pretrained(base_model_name)
base_model = AutoModelForCausalLM.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(base_model, adapter_model_name)

prompt = "Your messy telemetry and triage note string here..."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))