README.md

---
license: cc-by-sa-4.0
task_categories:
- text-generation
- token-classification
- feature-extraction
language:
- en
size_categories:
- 100K<n<1M
pretty_name: '​Stanza-2: Geometry-Aware WikiText'
---
# Dataset Card for Stanza-2

## Dataset Description
Stanza-2 is a structurally pristine, mathematically verified NLP dataset designed specifically for multi-task language modeling, custom tokenizer training, and mechanistic interpretability research. 

It is a rigorously modernized and annotated derivative of the `wikitext-2-raw-v1` corpus. By utilizing the Stanford NLP `Stanza` pipeline, every word in the corpus has been explicitly mapped to its grammatical, syntactic, and semantic function. Crucially, Stanza-2 preserves document geometry, explicitly labeling Markdown headers to support structure-aware neural architectures.

- **Curated by:** Jonathan R. Belanger (Exorobourii LLC)
- **Language:** English (`en`)
- **License:** CC-BY-SA-4.0
- **Total Rows:** 101,455 sentences (~2.46 Million Tokens)

## Dataset Structure
Stanza-2 abandons flat-text formatting in favor of **Parallel Arrays**. Each row in the dataset represents a single sentence. The linguistic features of that sentence are stored in perfectly aligned, equal-length arrays, guaranteeing 1:1 token-to-tag mapping.

### Schema
* `chunk_id` (int64): The positional ID of the chunk within the document stream.
* `sentence_id` (int64): The positional ID of the sentence within its chunk.
* `raw_text` (string): The cleaned, normalized raw string of the sentence.
* `is_header` (bool): `True` if the sentence is a structural document header.
* `section_level` (int64): The Markdown depth of the header (1-6). `0` if not a header.
* `tokens` (list[str]): The tokenized string sequence.
* `lemmas` (list[str]): The base morphological root of each token.
* `upos` (list[str]): Universal Part-of-Speech tags.
* `xpos` (list[str]): Treebank-specific Part-of-Speech tags.
* `head` (list[int64]): The 1-based index of the syntactic parent (Dependency Graph).
* `deprel` (list[str]): The syntactic dependency relation to the head token.
* `ner` (list[str]): Named Entity Recognition tags in explicit BIOES format.

## Methodology & Provenance

### 1. Cryptographic Ingestion
To prevent silent upstream updates from compromising downstream reproducibility, this dataset was built from a cryptographically verified snapshot of the `ggml-org/ci` raw mirror.
* **Source Archive:** `wikitext-2-raw-v1.zip`
* **SHA-256 Checksum:** `ef7edb566e3e2b2d31b29c1fdb0c89a4cc683597484c3dc2517919c615435a11`

### 2. The Normalization Ledger
The legacy WikiText corpus contains archaic spacing and tokenization artifacts. Prior to semantic enrichment, the text underwent strict, idempotent modernization passes to ensure sub-word tokenizers are not biased by historical formatting:
* **Hyphenation:** Legacy `@-@` artifacts were strictly mapped to standard hyphens (`-`).
* **Punctuation Alignment:** Floating terminal punctuation (e.g., `word , word`) and floating brackets were realigned to their preceding/succeeding semantic tokens. *Note: Vectorized backreferences were routed through standard Python CPU processing to bypass known `libcudf` regex injection vulnerabilities.*
* **Structural Preservation:** Legacy `= Header =` formats were mapped to standard Markdown (`# Header`) using strict descending-order regex (H6 down to H1) to prevent partial matching and preserve true document hierarchy.

### 3. Graph Integrity Protocol
Following the Stanza `depparse` enrichment, the resulting Parquet files were subjected to a microscopic mathematical audit. The Stanza-2 dataset guarantees 100% structural integrity:
1. **Dimensional Symmetry:** Every parallel array (`tokens`, `upos`, `ner`, etc.) within a row is guaranteed to be the exact same length.
2. **Root Singularity:** Every sentence possesses exactly one dependency root (`head == 0`).
3. **Graph Bounds:** No dependency head points to an index outside the bounds of the sentence. 
*Note: During the final Phase 4b integrity audit, 8 sentences out of ~101,463 across the training split violated graph bounds or root singularity due to extreme source fragmentation. These 8 rows were surgically dropped to preserve absolute dataset-wide mathematical validity.*

### 4. Structural Grammar Baseline
Analysis of the `wiki.train` split reveals exactly 451 unique `(UPOS, DepRel)` structural combinations across ~2.46 million tokens, demonstrating a highly rigid grammatical scaffold suitable for entropy reduction in custom tokenizer design.

## Usage
Because the dataset uses PyArrow-backed lists for parallel arrays, loading it into standard ML pipelines is highly efficient:

```python
import pandas as pd
df = pd.read_parquet("hf://datasets/EXOROBOURII/Stanza-Wikitext-2/wiki.train.enriched.parquet")

# Example: Accessing perfectly aligned tokens and their dependency relations
first_sentence_tokens = df.iloc[0]['tokens']
first_sentence_relations = df.iloc[0]['deprel']