Add docs

annotation_guidelines.md

@@ -0,0 +1,48 @@
+# 🇲🇲 အညွှန်းသတ်မှတ်ရန် လမ်းညွှန်ချက်
+
+## အညွှန်း Label များ
+
+### 1. Positive (အပြုသဘော) ✅
+- **လက္ခဏာ**: ရိုးသားသော ကျေးဇူးတင်မှု၊ မင်္ဂလာ
+- **ဥပမာ**: 
+  - "ကျေးဇူးပါ"
+  - "အရမ်းကျေးဇူးတင်ပါတယ်"
+  - "မင်္ဂလာပါ"
+
+### 2. Negative (အနှုတ်သဘော) ❌
+- **လက္ခဏာ**: မကျေနပ်ခြင်း၊ ဒေါသ၊ တင်းမာမှု
+- **ဥပမာ**:
+  - "ကျေးဇူးပါ" (သရော်သည်ဖြစ်နိုင်)
+  - "ဘာကြောင့်လည်းမသိ"
+
+### 3. Neutral (အလယ်အလတ်) ➖
+- **လက္ခဏာ**: သတင်းအချက်အလက်သာဖြစ်
+- **ဥပမာ**:
+  - "ကျေးဇူးပါ" (ရိုးရှင်းစွာပြောသည်)
+
+### 4. Sarcastic (သရော်ခြင်း) 🤨
+- **လက္ခဏာ**: အဓိပ္ပါယ်ကွေ့၊ သရော်ခြင်း
+- **ဥပမာ**:
+  - "ကျေးဇူးပါဗျာ" (အနိုင်ယှဉ်းသည်)
+
+## Intensity (အားပြိုင်မှု)
+
+| Level | Score | Description |
+|-------|-------|-------------|
+| Very Low | 0.1-0.2 | အားနည်းသော ဆိုင်းငံ့ |
+| Low | 0.3-0.4 | ပုံမှန် |
+| Medium | 0.5-0.6 | ပြင်းပြ |
+| High | 0.7-0.8 | အလွန်ပြင်း |
+| Very High | 0.9-1.0 | အထူးပြင်း |
+
+## Prosody (အသံ) ဂရုစိုက်ရန်
+
+- **Pitch (အသံအနိမ့်အမြင့်)**: မြင့် = စိတ်လှုပ်ရှား၊ နိမ့် = ငြိမ့်ညော
+- **Speed (အမြန်နှုန်း)**: မြန် = တင်းမာ၊ နှေး = ပျော်ပါ
+- **Pause (နားချိန်)**: ရှည် = စဉ်းစား၊ မရှိ = အလိုလိုသော
+
+## မှတ်စု
+
+1. Text နှင့် Audio prosody နှစ်ခုလုံးကို ကြည့်ပါ
+2. Context ကို သတိပြုပါ (အရင်စကားပြောချက်)
+3. သံသယဖြစ်ပါက "Confidence" နိမ့်စွာ သတ်မှတ်ပါ

api_reference.md

@@ -0,0 +1,127 @@
+# API Reference
+
+## FastAPI Endpoints
+
+### Health Check
+```
+GET /health
+```
+Response:
+```json
+{
+  "status": "healthy",
+  "model_loaded": true
+}
+```
+
+### Predict Sentiment
+```
+POST /predict
+```
+Request:
+```json
+{
+  "text": "ကျေးဇူးပါ",
+  "include_prosody": false
+}
+```
+Response:
+```json
+{
+  "text": "ကျေးဇူးပါ",
+  "sentiment": "positive",
+  "confidence": 0.95,
+  "probabilities": {
+    "negative": 0.01,
+    "neutral": 0.02,
+    "positive": 0.95,
+    "sarcastic": 0.02
+  }
+}
+```
+
+### Batch Predict
+```
+POST /predict_batch
+```
+Request:
+```json
+{
+  "texts": ["ကျေးဇူးပါ", "မကျေနပ်ပါဗျ"]
+}
+```
+
+## Python SDK
+
+### Installation
+```bash
+pip install myanmar-ghost
+```
+
+### Usage
+```python
+from myanmar_ghost import MyanmarGhost
+
+# Initialize
+model = MyanmarGhost()
+
+# Predict
+result = model.predict("ကျေးဇူးပါ")
+print(result.sentiment)  # "positive"
+
+# Batch predict
+results = model.predict_batch([
+    "ကျေးဇူးပါ",
+    "မကျေနပ်ပါ"
+])
+```
+
+### Advanced Usage
+
+#### XAI Explanations
+```python
+from myanmar_ghost.xai import SHAPExplainer
+
+explainer = SHAPExplainer(model)
+shap_values = explainer.explain("ကျေးဇူးပါ")
+explainer.visualize(shap_values)
+```
+
+#### Active Learning
+```python
+from myanmar_ghost.active_learning import UncertaintySampler
+
+sampler = UncertaintySampler(model)
+selected = sampler.select_samples(unlabeled_data, n_samples=100)
+```
+
+## CLI Commands
+
+```bash
+# Train model
+python -m src.models.train --train_data data/train.csv --output_dir outputs/models
+
+# Evaluate model
+python -m src.models.evaluate --model_path outputs/models/best_model.pt --data_path data/test.csv
+
+# Deploy
+bash scripts/deploy_model.sh outputs/models/best_model.pt
+```
+
+## Configuration
+
+### Environment Variables
+| Variable | Description | Default |
+|----------|-------------|---------|
+| MODEL_PATH | Path to model files | outputs/models |
+| HF_TOKEN | HuggingFace token | None |
+| DEVICE | cuda or cpu | cuda |
+
+### Model Config
+```yaml
+model:
+  name: myanmar_ghost
+  hidden_size: 768
+  num_layers: 12
+  dropout: 0.1
+```

model_card.md

@@ -0,0 +1,95 @@
+# Myanmar Ghost Model Card
+
+## 🏷️ Model Overview
+
+**Model Name**: Myanmar-Ghost-Instruct  
+**Model Type**: Text Classification (Sentiment Analysis)  
+**Language**: Myanmar (Burmese)  
+**Version**: 1.0.0  
+**Last Updated**: 2025
+
+## 📊 Model Description
+
+Myanmar Ghost is an advanced sentiment analysis model for Myanmar language that classifies text into 4 sentiment categories with multi-modal capability (audio + text).
+
+### Capabilities
+- Myanmar text sentiment classification
+- Multi-modal fusion (audio prosody + text)
+- Explainable AI (SHAP, LIME)
+- Privacy-preserving (Federated Learning ready)
+
+### Limitations
+- Best performance on formal Myanmar text
+- May struggle with heavy use of emoji/emoticons
+- Limited performance on code-mixed text
+
+## 📈 Training Data
+
+- **Source**: Myanmar speech datasets
+- **Size**: ~1M samples
+- **Splits**: 80% train, 10% validation, 10% test
+
+## ⚙️ Model Architecture
+
+```
+Transformer (BERT-based multilingual)
+├── Hidden Size: 768
+├── Layers: 12
+├── Heads: 12
+└── Classifier Head
+    └── 4-class output (negative, neutral, positive, sarcastic)
+```
+
+## 📉 Performance
+
+| Metric | Score |
+|--------|-------|
+| Accuracy | ~92% |
+| F1 (weighted) | ~91% |
+| F1 (macro) | ~89% |
+| Precision | ~91% |
+| Recall | ~91% |
+
+## 🔧 Usage
+
+### Python
+```python
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+
+model_name = "amkyawdev/Myanmar-Ghost-Instruct"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+
+# Predict
+text = "ကျေးဇူးပါ"
+inputs = tokenizer(text, return_tensors="pt")
+outputs = model(**inputs)
+```
+
+### API
+```bash
+curl -X POST http://localhost:8000/predict \
+  -H "Content-Type: application/json" \
+  -d '{"text": "ကျေးဇူးပါ"}'
+```
+
+## ⚠️ Ethical Considerations
+
+- Model trained on publicly available Myanmar data
+- No personally identifiable information used
+- Regular evaluation for bias
+
+## 📝 Citation
+
+```
+@software{myanmar_ghost,
+  title = {Myanmar Ghost},
+  author = {Aung Myo Kyaw},
+  url = {https://huggingface.co/amkyawdev/Myanmar-Ghost-Instruct},
+  year = {2025},
+}
+```
+
+## 🤝 License
+
+Apache 2.0