Upload README.md with huggingface_hub

README.md

@@ -1,24 +1,104 @@
+# vaovao_malagasy_sentiment_corpus
+
+# Dataset Card for Vaovao Malagasy Sentiment Corpus (VMSC)
+
+## Dataset Summary
+
+The **Vaovao Malagasy Sentiment Corpus (VMSC)** is the first publicly available, manually annotated sentiment dataset for the Malagasy language. It contains **5,041 sentences** from news articles published between 2022 and 2023, each labeled with binary sentiment (positive or negative).
+
+This dataset was created by native speakers using a high-quality annotation protocol, achieving **Cohen’s Kappa = 0.96**. VMSC enables research in sentiment analysis and other downstream NLP tasks for Malagasy, a low-resource language spoken by over 25 million people.
+
 ---
-dataset_info:
-  features:
-  - name: text
-    dtype: string
-  - name: label
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 578207.2763340607
-    num_examples: 4536
-  - name: validation
-    num_bytes: 64372.723665939295
-    num_examples: 505
-  download_size: 357827
-  dataset_size: 642580.0
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
+
+## Supported Tasks
+
+- Sentiment Analysis (binary classification)
+- Text Classification
+- Sentence-level Language Modeling
+- Benchmarking for Low-resource NLP
+
 ---
+
+## Languages
+
+- **Language**: Malagasy (`mg`)
+- **Variant**: Standard Malagasy
+
+---
+
+## Dataset Structure
+
+- **Format**: CSV
+- **Fields**:
+  - `text`: a sentence in Standard Malagasy
+  - `label`: sentiment annotation (`0` = negative, `1` = positive)
+- **Samples**: 5,041
+- **Sentence Length**:
+  - Minimum: 9 characters
+  - Average: 107.5 characters
+  - Maximum: 415 characters
+- **Class Distribution**:
+  - Positive: 2,616 (51.9%)
+  - Negative: 2,425 (48.1%)
+
+---
+
+## Annotation
+
+
+- **Annotators**: 3 native Malagasy speakers with academic backgrounds
+- **Annotation Protocol**:
+  - Detailed guideline with examples and decision rules
+  - Annotators labeled all examples independently
+  - Labels resolved via majority vote
+- **Agreement**:
+  - Unanimous (3/3): 97.04%
+  - Majority (2/3): 2.96%
+  - No agreement (0/3): excluded
+- **Cohen’s Kappa**: 0.96
+
+---
+
+## Source Data
+
+- **Collection Period**: Late 2022 – 2023
+- **Source Type**: Public Malagasy news websites
+- **Domains**:
+  - Politics
+  - Economy
+  - Society
+  - Culture and Media
+  - Environment and Agriculture
+- **Collection Method**:
+  - Rule-based sentence segmentation
+  - Manual cleaning and normalization
+
+---
+
+## Benchmark Results
+
+A wide range of machine learning and transformer models were benchmarked.
+
+| Model           | Accuracy | F1-score |
+|----------------|----------|----------|
+| Afro-XLM-R     | 0.7980   | 0.8111   |
+| XLM-R          | 0.7683   | 0.7636   |
+| mBERT          | 0.7603   | 0.7632   |
+| DistilBERT     | 0.7445   | 0.7455   |
+| Naive Bayes    | 0.7604   | 0.7695   |
+| SVM            | 0.7584   | 0.7645   |
+| Logistic Regression | 0.7505 | 0.7605 |
+| BiLSTM         | 0.7470   | 0.6861   |
+| BiGRU          | 0.7443   | 0.6861   |
+| CNN            | 0.6936   | 0.6861   |
+
+**Afro-XLM-R** outperformed all models in terms of both accuracy and F1-score.
+---
+
+## How to Use
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset('Lo-Renz-O/vaovao_malagasy_sentiment_corpus')
+print(dataset['train'][0])