README.md

metadata

license: apache-2.0
language:
  - ne
  - en
metrics:
  - accuracy
  - f1
  - precision
  - recall
base_model: sentence-transformers/all-MiniLM-L6-v2
new_version: 1.0.0
pipeline_tag: text-classification
library_name: scikit-learn
tags:
  - hybrid-model
  - logistic-regression
  - sentence-transformers
  - sbert
  - ne-en
  - rule-based
  - text-priority
  - low-resource-nlp
  - multilingual
  - civictech
  - complaint-triage
  - emergency-detection
eval_results:
  - task:
      type: text-classification
      name: Priority Detection (Nepali + English)
    dataset:
      name: priority_clean.csv (custom)
      type: csv
      size: 266 samples
    metrics:
      accuracy: 0.725
      f1_macro: 0.72
      precision_macro: 0.73
      recall_macro: 0.73
      per_class:
        HIGH:
          precision: 0.73
          recall: 0.66
          f1: 0.69
        MEDIUM:
          precision: 0.74
          recall: 0.8
          f1: 0.76
        LOW:
          precision: 0.71
          recall: 0.72
          f1: 0.71

Priority Classification Model (Nepali + English Hybrid)

Model Overview

This model automatically classifies citizen complaints or service requests into priority levels — HIGH, MEDIUM, or LOW — based on the urgency and nature of the text. It supports both Nepali and English inputs and uses a hybrid ML + rule-based approach to ensure robustness, especially on small datasets.

---

Model Architecture

Component	Description
Embedder	sentence-transformers/all-MiniLM-L6-v2
Classifier	Logistic Regression (multiclass, balanced weights)
Rule-based Layer	Keyword-based fallback for urgency terms in Nepali and English
Features	SBERT embeddings + priority keyword preservation
Hybrid Inference	Combines ML prediction confidence with rules for safer decisions

---

Training Summary

Metric	Value
Total raw samples	266
After preprocessing & augmentation	594
Train/Test Split	445 / 149
Embedding Dimension	384
Classes	HIGH, MEDIUM, LOW
Test Accuracy	72.5%
Macro F1-score	0.72

Label Distribution (After Normalization)

Label	Count
HIGH	203
MEDIUM	29
LOW	34

Label Distribution (After Augmentation)

Label	Count
HIGH	200
MEDIUM	194
LOW	200

---

Classification Report

Class	Precision	Recall	F1	Support
HIGH	0.73	0.66	0.69	50
MEDIUM	0.74	0.80	0.76	49
LOW	0.71	0.72	0.71	50
Overall Accuracy			0.725	149

Performance is acceptable (≥70%) given dataset size. The model performs best on clearly marked “urgent/emergency” cases and slightly lower on borderline MEDIUM cases.

---

Inference (Usage)

Using the model directly (ML only or Hybrid)

from huggingface_hub import hf_hub_download
import joblib
from priority_det import Embedder, predict_priority

# Download the model
model_path = hf_hub_download(repo_id="your-username/priority-classifier", filename="classifier.joblib")

# Load the classifier
bundle = joblib.load(model_path)
clf = bundle["clf"]
label_map = bundle["label_map"]

# Initialize the embedder
embedder = Embedder()

# Predict
text = "पानी आपूर्ति बन्द छ। तत्काल समाधान चाहिन्छ।"
result = predict_priority(text, embedder, clf, label_map, use_hybrid=True)
print(result)