Ash11/ntina-text-dataset-v0

Ntina Text Dataset V0

Ntina Text Dataset V0 is the text companion to Ntina Voice Dataset V0. It packages the project's literacy assessment material, bilingual tutor scaffolding, elicitation prompts, and dialogue-act specification as separate Hugging Face configurations.

The code-switched scaffolding is synthetic model output followed by human review. tools/generate-scaffolding.js records the generation model identifier as claude-sonnet-4-6. Reviewers evaluated every draft as keep, fix, or reject; deployable final_text contains the kept line or supplied repair.

This is a public V0 dataset. Standard Kiswahili content was verified by two Swahili speakers, one from Tanzania and one from Congo. Standard Yoruba content was approved and verified by two Yoruba speakers and a linguist.

Published repository: Ash11/ntina-text-dataset-v0.

Configurations

Configuration	Rows	Contents	Current status
scaffolding	176	88 Swahili–English and 88 Yoruba–English tutor utterances	Approved keep and repaired fix rows
instrument	156	Standard Kiswahili and Standard Yoruba TaRL reading items	Approved
prompts	30	Ten Set B prompts in English, Standard Kiswahili, and Standard Yoruba	Approved/reference text
dialogue_acts	88	Language-neutral pedagogical functions and triggers	Developer-authored specification

Total: 450 rows across four configurations. The experimental Tanzanian adaptation pilot is intentionally excluded from this main dataset.

Validation

All four configurations load successfully with the Hugging Face datasets library. The exported CSV files are UTF-8, contain no empty records, and have no duplicate primary keys within a language or language-pair configuration. The scaffolding export contains 54 normalized keep decisions and 122 fix decisions. No unrepaired rejection is included.

Scaffolding decisions

The scaffolding configuration retains both the raw reviewer verdict and a normalized decision:

• keep: final_text is the original draft.
• fix: the reviewer supplied a replacement, and final_text is that replacement.
• reject: no usable replacement was supplied, so the row is excluded from the deployable configuration.

This implements the project's review rule that a row with a good replacement is a fix, even when the raw verdict begins with the word “reject.” The raw verdict remains available in review_decision for auditability.

Linguist evaluation findings

The Yoruba linguist review identified wrong code-switch words or labels, incorrect tone marking, unnecessary word additions, and generated lines with no actual code-switch. The Swahili review found repeated translation-like openers—especially overuse of Sawa as a direct equivalent of English “Okay”—and other phrasing that preserved the English intent without sounding like natural bilingual classroom speech. These are qualitative observations, not measured corpus-wide error rates. The final dataset contains reviewed or repaired text rather than unfiltered drafts.

Full methodology and examples are included in the synthetic-data evaluation.

Yoruba Unicode recovery

The linguist's reviewed workbook preserved the Yoruba subdots and tone marks, while an earlier CSV export replaced some of them with literal question marks. The working yo-en.csv was therefore rebuilt from the intact UTF-8 draft by copying only the workbook's verdict and corrected_text columns. All 88 dialogue-act IDs and all ten non-review columns were checked for exact agreement before replacement. The damaged export is no longer used.

The recovery is reproducible with tools/recover-yoruba-scaffolding.py. The original .xlsx workbook remains as review provenance; users do not need it to load this Hugging Face export.

For future reviews, this recovery script can safely transfer the linguist's review fields from .xlsx to UTF-8 CSV without losing Yoruba subdots or tone marks. It validates the row IDs and unchanged source columns before creating the CSV, avoiding the character damage caused by some spreadsheet CSV exports.

Recommended use

• Use normalized_decision and non-empty final_text when selecting reviewed scaffolding.
• Retain act_id to join scaffolding rows to the dialogue_acts configuration.
• Use the reviewed_by, review_status, and provenance fields when auditing the approved instrument rows.

Sources and licensing

The dataset uses custom terms (license: other) because its reading material has mixed provenance. Row-level source and attribution details are retained in the instrument configuration.

Assessment methodology

• TaRL Assessment Tools provided the overall reading-level structure.
• The EGRA Toolkit, second edition and EGRA Applications and Interventions informed the use of Kiswahili syllables, including the ka, ke, ki, ko, ku pattern; familiar Grade 1–3 vocabulary; frequency-based item selection; and short stories with literal and inferential questions.
• The recommended EGRA frequency procedure was not completed: the project did not sample 20–30 textbook pages or calculate national textbook frequencies. The 30 syllables and 40 words in each language were author-created and later speaker/linguist verified, but they are not curriculum-frequency samples.
• The EPDC UWEZO/ASER compilation was located, but its Kenya and Tanzania test-item pages were scanned images and were not visually read during generation. No item text was copied or adapted from those test pages.

Narrative material actually used

African Storybook supplied the principal source material for paragraphs and stories:

• Swahili: Sare za shule (paragraph 1), Moto (paragraph 2), Mwanamume mrefu (story 1), and Tom muuza ndizi (story 2).
• Yoruba: African Storybook/Storybooks Nigeria material includes Ọkùnrin gíga púpọ̀ kan. One Yoruba story and much of the short-item material are author-created, as identified in the row-level source fields.
• Storybooks Tanzania states that its African Storybook collection was repurposed for Tanzanian audiences. Its Moto page credits Matteo E. Mwita as translator and Lauwo George as reader.
• Storybooks Nigeria states that its collection was repurposed for Nigerian audiences. The Yoruba Ọkùnrin gíga púpọ̀ kan page credits Taiwo Ẹhinẹni as translator.

Most adapted African Storybook material is CC BY 3.0 or CC BY 4.0. The adapted Sare za shule item is CC BY-NC 3.0 and retains a non-commercial restriction.

Sources identified later, not used to generate the 70 short Swahili items

These sources are valuable for future country-alignment work but were not the original basis of the author-created Swahili syllables and words:

• Tanzania Standard I–II syllabus
• TIE reading-teaching module
• BAKITA publications
• KICD Lower Primary curriculum

These limitations do not override the human variety verification: this release labels the content Standard Kiswahili and Standard Yoruba. They do mean the short items must not be described as copied from, statistically selected from, or officially approved by a national curriculum authority.

Public availability does not waive source attribution, non-commercial, or other underlying requirements. Contact ashuza1411@gmail.com with licensing, reuse, or correction questions.

Rebuilding

From the project root:

node tools/prepare-text-dataset-v0.js

The main package builder is scripts/prepare-text-dataset-v0.js. The approval-recording and optional .xlsx recovery implementations are scripts/approve-main-text-v0.js and scripts/recover-yoruba-scaffolding.py. The recovery script is only needed when the reviewed workbook is present; the main Hugging Face build reads the already recovered UTF-8 yo-en.csv.

Hugging Face repository: Ash11/ntina-text-dataset-v0.