Create README.md

README.md

@@ -0,0 +1,125 @@
+---
+language:
+- ku
+size_categories:
+- 1B<n<10B
+---
+
+# Dataset Card for KurCorpus 2B
+
+This dataset card documents KurCorpus 2B, a multidialectal Kurdish text corpus intended for large-scale language modeling and downstream NLP research.
+
+This dataset card aims to be a base template for new datasets. It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/datasetcard_template.md?plain=1).
+
+## Dataset Details
+
+### Dataset Description
+
+KurCorpus 2B is a large-scale Kurdish corpus with more than 2 billion tokens spanning the three major dialects of Kurdish: Sorani (ckb), Badini/Kurmanji (kmr), and Hawrami/Gorani (hac). It is released to advance research on low-resource languages and to support training of multidialectal language models such as KurBERT 2B.
+
+- **Curated by:** Karwan Mahdi Rawf, Abdullah Abdullah, Amanj Hussein, Haukar Mohammed
+- **Funded by [optional]:** More Information Needed
+- **Shared by [optional]:** University of Halabja
+- **Language(s) (NLP):** Kurdish Sorani (ckb), Kurmanji/Badini (kmr), Hawrami/Gorani (hac)
+- **License:** CC BY 4.0
+
+### Dataset Sources [optional]
+
+- **Repository:** https://huggingface.co/datasets/abdulhade/Kurdishcorpus
+- **Paper [optional]:** More Information Needed
+- **Demo [optional]:** More Information Needed
+- **External record:** Mendeley Data DOI 10.17632/fb5xhhn6m5.1
+
+## Uses
+
+### Direct Use
+
+- Pretraining and finetuning of Kurdish language models
+- Text classification, sentiment analysis, and topic modeling in Kurdish
+- Token-level tasks such as NER and POS tagging after suitable labeling
+- Cross-dialect adaptation and transfer learning studies
+- Machine translation experiments involving Kurdish dialects
+- Information extraction, knowledge graph construction, and QA pipelines
+
+### Out-of-Scope Use
+
+- High-stakes decision making without human oversight, such as medical, legal, or safety-critical advice
+- De-anonymization, profiling, or identification of individuals
+- Generating abusive or harmful content or targeted harassment
+- Any use that violates source site terms or applicable laws
+
+## Dataset Structure
+
+The dataset is provided as large compressed archives or text shards suitable for large-scale training. Common layouts:
+- One or more archives (e.g., `.7z` or `.rar`) containing normalized text files
+- Alternatively, multiple `.txt` or `.txt.gz` shards for streaming
+
+If consumed as text shards, each sample is a line of UTF-8 text with a single field:
+- **text**: string
+
+No official train/validation/test splits are provided. Users should create splits appropriate to their tasks and ensure no domain leakage.
+
+## Dataset Creation
+
+### Curation Rationale
+
+Kurdish is historically under-resourced with substantial dialectal variation and orthographic inconsistency. KurCorpus 2B aims to provide a large, cleaned, and consistent text resource to enable robust Kurdish NLP, including dialect-aware modeling and cross-dialect transfer.
+
+### Source Data
+
+Sources include publicly accessible Kurdish text from:
+- News websites
+- Social media platforms such as Facebook and Telegram
+- Digitized literary texts and online publications
+
+#### Data Collection and Processing
+
+A dialect-aware normalization and cleaning pipeline was applied:
+- Unicode normalization (NFKC) for consistent code points
+- Orthographic correction for Arabic-script variants (e.g., ك -> ک, ي -> ی)
+- Noise removal of emojis, URLs, emails, and HTML with placeholders like [URL], [EMAIL]
+- Whitespace and punctuation normalization
+- Dialect-aware stopword filtering to reduce noise while preserving structure
+- Sentence boundary and token-level cleaning to improve modeling readiness
+
+Tools included standard Python text-processing libraries and custom normalization scripts. Exact heuristics and dictionaries can be expanded in future releases.
+
+#### Who are the source data producers?
+
+Original content was produced by Kurdish news outlets, social media users, and authors of literary or online texts in Kurdish. No private or non-public sources were intentionally collected.
+
+### Annotations [optional]
+
+No manual annotations are included in this release.
+
+#### Annotation process
+
+Not applicable.
+
+#### Who are the annotators?
+
+Not applicable.
+
+#### Personal and Sensitive Information
+
+The corpus aggregates public web text and may include names of public figures or references to events. Emails and URLs were replaced with placeholders where detected, but residual personal or sensitive content may remain. Users must comply with the license and with applicable data protection laws in their jurisdictions.
+
+## Bias, Risks, and Limitations
+
+- **Domain bias:** Over-representation of news and social media may skew styles and topics.
+- **Dialect balance:** While multidialectal, some dialects or sub-varieties may be underrepresented.
+- **Orthography and script:** Orthographic ambiguity and mixed conventions can persist despite normalization.
+- **Safety:** Content may include sensitive or offensive text typical of web corpora.
+
+### Recommendations
+
+- Apply task-specific filtering, deduplication, and domain balancing where needed.
+- Use human-in-the-loop review for safety-critical or production deployments.
+- Report dialectal coverage and distribution when publishing results.
+- For benchmarking, create clean, transparent splits and avoid overlap with evaluation sets.
+
+## Citation [optional]
+
+If you use KurCorpus 2B, please cite the Mendeley Data record:
+
+**BibTeX:**