Thermostatic/rosettia-chanka-data

RosettIA Chanka Quechua — Judicial Parallel Data

Spanish ↔ Chanka / Ayacucho Quechua (quy) parallel data for the RosettIA project.

⚠️ Scope of this public release (please read)

This repository only hosts the data we can redistribute cleanly — the Spanish–Chanka pairs and glossary extracted from a Peruvian Ministry of Culture manual that explicitly permits reproduction with attribution.

Our models were also trained on broad, synthetic, and reinforcement-learning corpora built from third-party sources with restrictive, viral, or unconfirmed licenses (JW300-derived AmericasNLP data, SomosNLP, FLORES-200, a published dictionary, web-mined text, C4). We do not redistribute any of that here. It is fully documented below, with source links, licenses, and the scripts to reconstruct it from the original providers under their own terms. See "Data used but NOT redistributed here."

What is included

All files are derived from a single, explicitly-redistributable source (see License & credits):

File	Rows	Content
clean_chanka/manual_quechua_chanka_parallel_training_ready_augmented.parquet	1,055	Reviewed Spanish–Chanka judicial/administrative pairs (slash-alternatives split)
clean_chanka/manual_quechua_chanka_parallel_reviewed.parquet	1,012	Reviewed extraction rows with decisions/flags
clean_chanka/manual_quechua_chanka_parallel_alternative_splits.parquet	81	Rows created from reviewed slash-alternative pairs
clean_chanka/manual_quechua_chanka_glossary_entries.parquet	423	Glossary entries from the manual's vocabulary sections
clean_chanka/manual_quechua_chanka_glossary_simple_terms.parquet	219	Simple glossary term pairs

Columns: reviewed_spanish, reviewed_chanka_quechua (+ review/flag metadata). Domain: judicial / administration of justice (not general-domain Chanka). Provenance notes are in metadata/.

License & credits

License: CC-BY-4.0 (attribution required).

All included data is extracted from:

Ardito Vega, W. (2014). Manual para el empleo del Quechua Chanka en la administración de justicia. Ministerio de Cultura del Perú, 1ª ed., Lima, abril 2014. ISBN 978-612-4126-20-8. Source PDF: https://repositorio.cultura.gob.pe/handle/CULTURA/51

The manual states (p. 3): "Se permite la reproducción de esta obra siempre y cuando se cite la fuente." (Reproduction is permitted provided the source is cited.) We therefore release the extracted/derived parallel pairs and glossary (not the PDF) under CC-BY-4.0, and require attribution to:

• Content author: Wilfredo Ardito Vega
• Collaborators: Eber Llacctarimay Quispe, Cinthya Palomino Córdova, Juan Galiano Román, Apolinario Ciriaco Saldívar Bolívar, Clodomiro Landeo Lagos, Salvador Alvarado Tovar, Carlos Andrés Vera Vásquez
• Editor: Mónica Hidalgo Cornejo · Publisher: Ministerio de Cultura del Perú

If you use this data, please cite the manual above and the RosettIA project.

Data used but NOT redistributed here

Our models were trained on additional corpora that we cannot redistribute (prohibited, viral, non-commercial, or unconfirmed licenses). They are not in this repository. To reproduce our training data, fetch each from its original provider under its own terms using the build scripts in the GitHub repo (scripts/).

Source (used for)	Link	License / status	Why not hosted here
AmericasNLP 2024 ST1 quy (broad + NLLB + RL data) — JW300-derived	github.com/AmericasNLP/americasnlp2024; JW300 paper	Not redistributable — JW300 was distributed without the copyright holder's (Watch Tower) permission and withdrawn	Prohibited
SomosNLP spanish-to-quechua (broad SFT)	HF dataset	No license stated (mixes Tatoeba CC-BY, biblical, AmericasNLP)	No granted right
FLORES-200 (inside NLLB corpus)	facebook/flores	CC-BY-SA-4.0 (viral share-alike; provider discourages re-hosting eval data)	Avoid share-alike entanglement + eval-integrity
Diccionario Básico Quechua Chanka (Benito Zuasnabar, 2018)	published dictionary	Copyrighted / unconfirmed	No open license
nouman-10/MT-SharedTask incl. Le Petit Prince (v31/v32 experiments)	github	No repo license; Little Prince still under copyright	Not redistributable
RunaSimi.de Ayacucho vocab (v31 experiment)	runasimi.de	© Philip Jacobs — non-commercial + attribution	Non-commercial
Synthetic targets (NLLB distillation): Spanish from C4 / OPUS → quy by our model	allenai/c4 (ODC-BY), OPUS	quy = our model output; Spanish carries C4 ODC-BY + Common Crawl ToS	Avoid hosting C4-derived web text

(Note: the experimental v31/v32, v34-normalizer, broad-ablation, and eval_mbr/ result files are also not part of this release — this dataset is scoped to the data actually used, and only the cleanly-licensed part of it.)

Authors & contributions

A two-person SomosNLP hackathon project:

• Estefanía Espinosa Fernández — data curation (including the reviewed extraction and cleaning of the Chanka pairs and glossary released here), and the initial Qwen3.5 LoRA experiments (DoRA / rsLoRA / LoRA comparison, data mixes).
• Irving Ernesto Quezada Ramírez (irvingernesto.com) — the subsequent modeling and release work (synthetic distillation, the NLLB pipeline, GSPO RL, decoding/ensembling, evaluation, and packaging).

The project was a close collaboration; both contributions were essential to the result. (Source attribution for the underlying Manual is in License & credits above.)

Citation

@misc{rosettia_chanka_data,
  title  = {RosettIA Chanka Quechua — Judicial Parallel Data},
  author = {Espinosa Fern\'andez, Estefan\'ia and Quezada Ram\'irez, Irving Ernesto},
  note   = {Derived from: Ardito Vega, W. (2014), Manual para el empleo del Quechua Chanka
            en la administración de justicia, Ministerio de Cultura del Perú,
            ISBN 978-612-4126-20-8},
  url    = {https://github.com/Sekinal/rosettia-chanka},
  year   = {2026}
}