Hugging Face's logo

kmanikandan
/
atman-healthai-medical-lora

PEFT
Safetensors
#healthcare
#medical
# Atman HealthAI Medical LoRA
#hackathon
#lora
Model card Files
xet
 Community

Atman HealthAI Medical LoRA

Model Overview

Atman HealthAI Medical LoRA is a healthcare-focused LoRA adapter fine-tuned for medical report explanation, symptom interpretation, risk-aware health guidance, and specialist recommendation. It was developed as part of the Adaption AutoScientist Challenge × HackIndia to make healthcare information easier to understand, especially for users who may not be comfortable reading medical terminology or English-only medical content.

This repository contains the LoRA adapter weights for a model fine-tuned on top of meta-llama/Llama-4-Scout-17B-16E-Instruct. It is designed to explain common lab values and symptom descriptions in simple language and return structured educational healthcare guidance.

Model Details

Base Model

  • Base model: meta-llama/Llama-4-Scout-17B-16E-Instruct

Fine-Tuned Model

  • Model name: adaption_medical_lab_result_advice
  • Project name: Atman HealthAI
  • Training platform: Adaption AutoScientist
  • Training method: Supervised Fine-Tuning (SFT)
  • Fine-tuning type: LoRA / PEFT

Developer

  • Developed by: Manikandan K / Team AtmanAI
  • Challenge: Adaption AutoScientist Challenge × HackIndia

Languages

Primary language is English, with training data adapted for multilingual healthcare assistance across major Indian languages including:

  • Tamil
  • Hindi
  • Telugu
  • Kannada
  • Malayalam
  • Bengali

What This Model Does

This model is designed to help with:

  • Medical report explanation Example: interpreting values like hemoglobin, blood glucose, WBC count, TSH, LDL, creatinine, etc.

  • Symptom interpretation Example: “fever and cough”, “fatigue and dizziness”, “chest pain while walking”

  • Risk-aware health guidance The model structures responses with risk levels such as:

    • Self Care Information
    • See Doctor Soon
    • Urgent Medical Attention

  • Specialist recommendation Example: General Physician, Endocrinologist, Cardiologist, Dermatologist, etc.

  • Health insight generation It provides a simple, educational explanation of what the symptom or report may indicate.

Intended Output Format

The model was trained to produce structured healthcare responses in this format:

Explanation: ...
Risk Level: ...
Recommended Specialist: ...
Health Insight: ...
Disclaimer: ...

Example

Input

Hemoglobin: 9.8 g/dL

Output

Explanation: Your hemoglobin level is lower than the normal range and may suggest anemia, which can cause tiredness, weakness, or dizziness.
Risk Level: See Doctor Soon
Recommended Specialist: General Physician
Health Insight: Low hemoglobin can happen due to iron deficiency, blood loss, or other medical conditions, so follow-up testing may be helpful.
Disclaimer: This is for educational purposes only and not a diagnosis. Please consult a qualified medical professional.

Training Data

This model was trained on a healthcare instruction dataset built for medical report explanation and healthcare assistance.

Dataset coverage includes:

  • Lab result interpretation
  • Symptom explanation
  • Healthcare question answering
  • Medical terminology simplification
  • Specialist recommendation
  • Risk-aware educational guidance

Dataset format

The training dataset uses instruction-response pairs with two columns:

  • input
  • output

Linked dataset

This model is paired with the dataset repository:

Dataset: atman-healthai-medical-dataset

Adaptive Data Results

The dataset was improved using Adaption Adaptive Data before model training.

Dataset quality improvements

  • Original dataset quality score: 6.0
  • Adapted dataset quality score: 8.1
  • Grade improvement: C → B
  • Percentile improvement: 2.4 → 17.8

These improvements reflect better structure, consistency, and training suitability of the healthcare dataset.

AutoScientist Training Results

The model was trained using Adaption AutoScientist.

Win-rate improvement over baseline

  • Base model win rate: 34
  • Adapted model win rate: 66
  • Absolute improvement: +32 points

This indicates that the fine-tuned model performed substantially better than the baseline in Adaption’s evaluation workflow for the task.

Training Configuration

Fine-tuning setup

  • Training method: SFT
  • Training type: LoRA
  • Data format: chat

Hyperparameters

  • Epochs: 1
  • Learning rate: 0.0001
  • LoRA rank (r): 64
  • LoRA alpha: 128
  • LoRA dropout: 0
  • Batch size: max
  • Scheduler: cosine
  • Warmup ratio: 0.05
  • Weight decay: 0.015
  • Max gradient norm: 1
  • Trainable modules: all-linear
  • Train on inputs: false

How to Use

This repository contains LoRA adapter weights, not the full original base model. To use the model, load the adapter on top of the base model:

meta-llama/Llama-4-Scout-17B-16E-Instruct

The general workflow is:

  1. Load the base model
  2. Load the tokenizer
  3. Apply the LoRA adapter from this repository
  4. Use a healthcare prompt such as a lab value, symptom description, or short medical question

Intended Use

This model is intended for:

  • educational medical report explanation
  • symptom understanding assistance
  • healthcare literacy support
  • multilingual healthcare AI prototyping
  • demo applications for medical report interpretation

Out-of-Scope Use

This model is not intended for:

  • medical diagnosis
  • emergency decision making
  • treatment prescription
  • replacing doctors, specialists, or licensed healthcare professionals
  • use as a sole source of medical advice

Risks, Biases, and Limitations

  • The model is trained for educational healthcare assistance, not clinical diagnosis.
  • It may oversimplify medical conditions or fail to capture full clinical context.
  • Some lab values require patient history, age, sex, medications, and co-existing conditions for proper interpretation.
  • The model should not be used for emergency care decisions.
  • Multilingual outputs may vary in quality depending on the language and phrasing of the prompt.
  • Rare diseases, complex medical cases, and specialist-only scenarios may not be handled reliably.

Safety Disclaimer

This model is for educational and informational purposes only. It is not a medical diagnosis system. Users should consult a qualified medical professional for diagnosis, treatment, or emergency care decisions.

Project Context

This model was created for the Adaption AutoScientist Challenge × HackIndia under the project:

Atman HealthAI

Problem addressed

Millions of people receive medical reports or experience symptoms but struggle to understand what they mean. This leads to confusion, anxiety, delayed action, and low healthcare accessibility—especially for users who prefer regional languages.

Atman HealthAI aims to make healthcare information more understandable by building a multilingual AI assistant that can:

  • explain reports in simple language
  • interpret symptoms
  • suggest the right specialist
  • improve health literacy across Indian languages

Authors

  • Manikandan K
  • Team AtmanAI

Contact / Project Links

Add your final links here after publishing:

  • Dataset repo: Hugging Face dataset link
  • Model repo: Hugging Face model link
  • GitHub project repo: your HackIndia project GitHub
  • Demo: add once deployed
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kmanikandan/atman-healthai-medical-lora

Base model

meta-llama/Llama-4-Scout-17B-16E
Finetuned
meta-llama/Llama-4-Scout-17B-16E-Instruct
Adapter
(27)
this model