README.md

---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: category
    dtype: string
  - name: extended_answer
    sequence:
    - name: user
      dtype: string
    - name: answer
      dtype: string
  splits:
  - name: train
    num_bytes: 363151819
    num_examples: 54348
  download_size: 136276329
  dataset_size: 363151819
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# DATASET: **Kazakh administrative documents for RAG document QA.**

* **Structure.** Each item is a JSON object with:

  * `text`: the full Kazakh document body (biography or power-of-attorney).
  * `category`: document type label — e.g., **Өмірбаян** (autobiographical CV/biography) and **Сенімхат** (power of attorney) etc. In Kazakh admin usage, *Өмірбаян* is a concise, chronological personal record; *Сенімхат* is a written authorization to act on someone’s behalf.
  * `extended_answer`: list of `{user, answer}` QA pairs extractable from `text` (factoid fields like birth date/place, degrees, awards; or principals/children, validity period, addresses, etc.).

* **Scope.** Kazakh-language **administrative and personal records** with explicit slot-like facts (names, dates, institutions, addresses, phone numbers) and templated legal phrasing (e.g., “сенімхат … жарамды”, placeholders like `[күні]`). The two shown categories align with common Kazakh document genres: autobiographies for employment/education workflows and powers of attorney for representation/transport of minors. 

* **Usage.**

  * **Extraction & slot filling:** train/evaluate NER/IE for structured fields (person, DOB, place, degree, positions; principal/agent, children, validity window).
  * **Document QA / RAG:** `extended_answer` provides supervision for extractive/generative QA grounded in `text`.
  * **Template validation & completion:** detect missing placeholders (e.g., `[күні]`) and verify mandatory fields typical for *сенімхат*; learn document-type–specific consistency rules.
  * **Document classification:** use `category` to train classifiers distinguishing biography vs. authorization letters.
  * **Privacy stress-tests:** includes realistic PII (addresses, phone numbers) to test redaction or safe-answering policies (if needed).

**Notes.** No cross-file alignment is required; each JSON is self-contained: raw text + gold QA pairs enable end-to-end pipelines (parse → retrieve within doc → answer). The dataset is suitable for low-resource Kazakh NLP where domain conventions for *өмірбаян* and *сенімхат* are well-defined.