mannsaathi-symptom-classifier-large

Model Card for MannSaathi Symptom Classifier (Large)

<!-- Provide a quick summary of what the model is/does. -->

A multilingual symptom classification model fine-tuned to understand how people in India actually describe health concerns — in Hindi, English, and Hinglish. Built as the AI core of MannSaathi, an anonymous health companion designed to help people overcome fear and hesitation around seeking healthcare.

Model Details

Model Description

This model classifies free-text symptom descriptions into 22+ symptom categories and outputs a calibrated risk score. It was fine-tuned using LoRA (Low-Rank Adaptation) on top of xlm-roberta-large, specifically trained to handle code-mixed Hindi-English (Hinglish) input, which most general-purpose medical NLP models fail to interpret correctly.

The model intentionally avoids alarming or diagnostic language in its design philosophy — it is built to inform and reassure, not to diagnose.

• Developed by: Arsh Verma
• Model type: Multi-label text classification (fine-tuned with LoRA)
• Language(s): Hindi, English, Hinglish (code-mixed)
• License: MIT
• Finetuned from model: xlm-roberta-large

Model Sources

• Repository: https://github.com/ArshVermaGit/mannsaathi
• Demo: https://mannsaathi-mukk.onrender.com
• Video walkthrough: https://www.youtube.com/watch?v=c58bVmaaMQ0

Uses

Direct Use

This model is intended to power conversational symptom-checking experiences where users describe how they're feeling in natural, informal language. It returns likely symptom categories and a risk indicator to help guide users toward an appropriate next step (self-care info, telemedicine, or in-person care).

Out-of-Scope Use

This model is not a diagnostic tool and must not be used as a substitute for professional medical advice, diagnosis, or treatment. It should not be deployed in any clinical decision-making system without human oversight. It is not validated for emergency triage.

Bias, Risks, and Limitations

The model was trained on a dataset skewed toward Hindi-English code-mixed text common in North Indian conversational patterns, and may underperform on other Indian languages or regional dialects. As with any symptom-classification model, false negatives on serious symptoms are possible — this model should always be paired with clear messaging encouraging users to seek professional care when uncertain.

Recommendations

Users (and downstream developers) should be made aware of the risks, biases, and limitations of the model. Always pair model output with clear next-step guidance and never present results as a clinical diagnosis.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("ArshVerma/mannsaathi-symptom-classifier-large")
model = AutoModelForSequenceClassification.from_pretrained("ArshVerma/mannsaathi-symptom-classifier-large")

text = "sir dard hai aur neend nahi aa rahi 3 din se"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits)

print(probs)

Training Details

Training Data

The model was trained on a custom-curated dataset of Hindi, English, and Hinglish health-related text, combining publicly available medical Q&A sources with manually labeled conversational symptom descriptions reflecting how people in India actually talk about their health.

Training Procedure

Fine-tuning was performed using LoRA (Low-Rank Adaptation) on top of xlm-roberta-large, allowing efficient training on limited GPU resources without updating all base model parameters.

Training Hyperparameters

• Training regime: Mixed precision (fp16)
• Base model: xlm-roberta-large
• Fine-tuning method: LoRA
• Hardware: Kaggle free-tier T4 GPU

Evaluation

Testing Data, Factors & Metrics

The model was evaluated on a held-out validation split of the curated dataset, measuring multi-label classification accuracy and F1 score across all 22+ symptom categories.

Metrics

• Accuracy
• F1 Score (macro-averaged)

Environmental Impact

Training was performed on a single Kaggle T4 GPU instance over a limited number of epochs, resulting in minimal estimated carbon impact compared to large-scale pretraining runs.

• Hardware Type: NVIDIA T4 GPU (Kaggle free tier)
• Cloud Provider: Kaggle (Google Cloud-backed)
• Compute Region: Unknown

Technical Specifications

Model Architecture and Objective

XLM-RoBERTa-large backbone with a fine-tuned classification head (via LoRA adapters) for multi-label symptom category prediction.

Compute Infrastructure

Hardware

Kaggle free-tier NVIDIA T4 GPU

Software

• transformers
• peft (for LoRA)
• torch

Citation

If you use this model, please consider citing the project:

@misc{mannsaathi2026,
  author = {Arsh Verma},
  title = {MannSaathi Symptom Classifier (Large)},
  year = {2026},
  url = {https://huggingface.co/ArshVerma/mannsaathi-symptom-classifier-large}
}

Model Card Contact

Arsh Verma — GitHub · Project