Create README.md

README.md

@@ -0,0 +1,156 @@
+---
+dataset_name: pashto-sentiment
+language:
+  - ps
+task_categories:
+  - text-classification
+task_ids:
+  - sentiment-classification
+license: apache-2.0
+size_categories:
+  - 1k<n<10k
+annotations_creators:
+  - expert-generated
+source_datasets:
+  - original
+pretty_name: Pashto Sentiment Dataset
+---
+
+
+
+# **Pashto Sentiment Dataset — 2026 Edition**
+
+A high‑quality **Pashto Sentiment Classification** dataset designed for training and evaluating Pashto NLP models.  
+This dataset provides **clean**, **normalized**, and **manually‑verified** sentiment labels suitable for SFT, supervised classification, and benchmarking Pashto LLMs.
+
+---
+
+## **📦 Dataset Summary**
+
+The **Pashto Sentiment Dataset** contains short and medium‑length Pashto texts annotated with one of three sentiment classes:
+
+- **Positive**
+- **Negative**
+- **Neutral**
+
+Each sample is normalized, deduplicated, and formatted for direct use in machine‑learning pipelines.
+
+---
+
+## **📊 Dataset Structure**
+
+### **Features**
+| Field | Type | Description |
+|-------|-------|-------------|
+| `text` | string | Pashto sentence or short paragraph |
+| `label` | string | One of: `"positive"`, `"negative"`, `"neutral"` |
+
+### **Example**
+```json
+{
+  "text": "دا کار ډېر ښه ترسره شوی دی او زه پرې خوشاله یم",
+  "label": "positive"
+}
+```
+
+---
+
+## **🧹 Cleaning & Normalization Pipeline**
+
+The dataset was processed using the **Talanda NLP pipeline**, including:
+
+- Unicode normalization (NFKC)
+- Pashto‑specific orthographic cleanup
+- Removal of duplicate or near‑duplicate samples
+- Length filtering (min 5 chars, max 512 chars)
+- Sentiment consistency checks
+- LMDB‑based caching for deterministic extraction
+
+---
+
+## **📁 File Format**
+
+The dataset is provided in **JSONL** format:
+
+```
+pashto-sentiment/
+ 
+```
+
+Each line is a single JSON object containing `text` and `label`.
+
+---
+
+## **🚀 Usage**
+
+### **Load with HuggingFace Datasets**
+```python
+from datasets import load_dataset
+
+ds = load_dataset("nassimjp/pashto-sentiment")
+
+print(ds["train"][0])
+```
+
+### **Use in a Classification Model**
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+
+tokenizer = AutoTokenizer.from_pretrained("bert-base-multilingual-cased")
+model = AutoModelForSequenceClassification.from_pretrained(
+    "bert-base-multilingual-cased",
+    num_labels=3
+)
+
+inputs = tokenizer(ds["train"][0]["text"], return_tensors="pt")
+outputs = model(**inputs)
+```
+
+---
+
+## **📈 Intended Applications**
+
+- Pashto sentiment classification models  
+- SFT for Pashto LLMs  
+- Benchmarking Pashto text‑understanding tasks  
+- Domain‑specific sentiment analysis (news, social media, reviews)
+
+---
+
+## **⚠️ Limitations**
+
+- Sentiment labels are not multi‑class (no fine‑grained emotions)
+- Dataset focuses on general‑domain text; domain‑specific corpora may require additional tuning
+- Some ambiguous sentences may have borderline sentiment
+
+---
+
+## **📜 License**
+
+This dataset is released under the **Apache 2.0 License**.
+
+---
+
+## **🤝 Citation**
+
+If you use this dataset in your research, please cite:
+
+```
+@dataset{nassimjp_pashto_sentiment_2026,
+  author       = {Nassim},
+  title        = {Pashto Sentiment Dataset},
+  year         = {2026},
+  publisher    = {HuggingFace Datasets},
+  url          = {https://huggingface.co/datasets/nassimjp/pashto-sentiment}
+}
+```
+
+---
+
+## **🔗 Related Datasets**
+
+- **PashtoNanoChat**  
+- **Pashto-SFT-263k**  
+- **Pashto Wheat SFT 14k**  
+
+---