--- language: - qu - es license: other license_name: mixed-apache-mit-gpl pretty_name: Kichwa–Spanish Parallel Lexicon & Instructions (Napo / Amazonian Kichwa) task_categories: - translation - text-generation multilinguality: - multilingual size_categories: - 10K ⚠️ **The Kichwa has not been validated by a fluent speaker.** Treat this as a cleaned aggregation of an > existing source, not an authoritative reference. See [Limitations](#limitations--caveats). ## What's inside | Config | Rows | License | Contents | |---|---|---|---| | `lexicon` *(default)* | 6,313 | Apache-2.0 + MIT | Parallel pairs (headwords **and** example sentences). 6,287 flagged `ok`, 26 `review`. | | `instructions` | 6,354 | MIT | Chat-formatted Q&A pairs (instruction tuning), derived from the same lexicon. | | `lexicon_gpl` | 7,543 | GPL-3.0 | Parallel pairs from the GPL-licensed source, kept **separate** (see [Licensing](#licensing)). 7,542 `ok`, 1 `review`. | Across both lexicon tiers there are **13,856** parallel pairs. The two tiers overlap (they share an upstream dictionary) but are not identical: the permissive tier has richer multi-gloss headwords and more example sentences; the GPL tier has broader single-gloss coverage. They are kept apart purely for licensing reasons, not because the content is unrelated. ## How to load ```python from datasets import load_dataset # Default: permissive parallel lexicon (Apache-2.0 + MIT) lex = load_dataset("aimeri/kichwa-spanish", "lexicon", split="train") # Clean rows only (drop the flagged ones) lex_clean = lex.filter(lambda r: r["quality_flag"] == "ok") # Instruction / chat data (MIT) inst = load_dataset("aimeri/kichwa-spanish", "instructions", split="train") # GPL lexicon — only pull this in if your project is GPL-compatible gpl = load_dataset("aimeri/kichwa-spanish", "lexicon_gpl", split="train") ``` ## Data fields **`lexicon` and `lexicon_gpl`:** | Field | Type | Description | |---|---|---| | `kichwa` | string | Kichwa side (headword or example sentence). | | `spanish` | string | Spanish side (gloss or translation). | | `entry_type` | string | `headword` (1–2 token term) or `example` (≥3 token sentence). Heuristic. | | `quality_flag` | string | `ok`, or `review` if the Kichwa field reads as net-Spanish (likely misaligned in source). | | `source` | string | Origin repo(s). `+`-joined when a deduplicated pair appeared in more than one source. | | `license` | string | Per-row license. `+`-joined when merged across sources. | **`instructions`:** | Field | Type | Description | |---|---|---| | `messages` | list | Chat turns: `[{"role": "user"/"assistant", "content": "..."}]`. | | `source` | string | Origin repo. | | `license` | string | `mit`. | ## Provenance & credits ### Upstream source dictionary All three contributing datasets trace back to the **_Diccionario Kichwa–Castellano_**, an Amazonian (Napo) Kichwa ↔ Spanish dictionary produced by **UNICEF** and Ecuador's **DINEIB** (Dirección Nacional de Educación Intercultural Bilingüe), ISBN 978-92-806-4082-3, published for free educational distribution (*"MANA KATUNACHU — DISTRIBUCIÓN GRATUITA"*). A digitized copy is hosted in the [FLACSO Andes digital library](https://repositorio.flacsoandes.edu.ec/items/da4f1fd3-1393-4cf6-a746-c02e1026b1e4). The dictionary's example sentences (e.g. river-Napo flora and fauna, *bocachico*, *capirona*, *ayawaska*) place it firmly in the **Napo Lowland** variety. The linguistic and cultural content originates with the Kichwa communities and the educators who compiled it; please respect the spirit of that work. ### Contributing Hub datasets (data collectors / publishers) This repo would not exist without three people who did the work of digitizing and publishing the data: - **[Carlos89apc/TraductorES_Kichwa](https://huggingface.co/datasets/Carlos89apc/TraductorES_Kichwa)** — GPL-3.0. ~8K tab-separated Spanish→Kichwa pairs (broadest coverage). → `lexicon_gpl`. - **[hcurbelo/Kichwa_ES_001](https://huggingface.co/datasets/hcurbelo/Kichwa_ES_001)** — Apache-2.0. Headwords + example sentences with richer multi-synonym glosses. → `lexicon`. - **[hcurbelohy/ES_Kichwa](https://huggingface.co/datasets/hcurbelohy/ES_Kichwa)** — MIT. Instruction-formatted Q&A derived from the same lexicon. → `instructions` (+ mined pairs into `lexicon`). ## Licensing This repo intentionally **splits content by license** so the permissive core stays clean: - **Permissive configs** — `lexicon` (Apache-2.0 + MIT) and `instructions` (MIT). Safe to use in most projects. Every row carries its own `license` field so you can filter further if needed. - **Copyleft config** — `lexicon_gpl` is **GPL-3.0** and is kept in its own config. If your project can't take GPL, simply don't load `lexicon_gpl`; nothing in the permissive configs depends on it. The repository-level license is marked `other` because the configs carry different licenses; the authoritative per-item license is the `license` column and the per-config notes above. Licenses are inherited from the original uploaders — the underlying dictionary is itself a free-distribution UNICEF/ DINEIB educational publication, which is acknowledged but does not by itself grant an SPDX license. ## Cleaning & normalization What was done to produce this repo from the three sources: - **Correct per-source parsing.** `hcurbelo`'s real data (a single `KICHWA###SPANISH` column that the Hub viewer had collapsed into one field) was split properly; `Carlos89apc`'s tab-separated pairs were parsed by direction; `hcurbelohy`'s JSON-in-CSV `messages` column (which caused the original viewer cast error) was decoded into real chat structures, and its templated Q&A was additionally mined back into parallel pairs. - **Language-orientation repair.** The sources mix column order, leaving a few percent of rows with Spanish where Kichwa should be (and vice-versa). Each row was re-oriented by a Kichwa-vs-Spanish scorer (Kichwa morphology and `sh`/`w`/`k` density vs Spanish function words, verb endings, `qu`/`gu`, and accents), with source order as a tie-break for ambiguous bare nouns. ~280 rows were corrected. - **Normalization.** Whitespace collapsed, edge artifact punctuation (`¡¿!.?`) stripped, all-caps headwords lowercased. - **Deduplication.** Exact duplicate `(kichwa, spanish)` pairs removed within each license tier; provenance merged (`source`/`license` joined with `+`). - **Quality flagging.** Rows whose Kichwa field still reads as net-Spanish — genuine source misalignments, not parsing errors — are marked `quality_flag = "review"` rather than silently shipped or deleted. ## Limitations & caveats - **Not speaker-validated.** No fluent Kichwa speaker has reviewed correctness. Where the two source dictionaries disagree on a word, both readings may be present across the tiers. - **Spanish-side OCR/wrapping artifacts.** The Spanish text carries spurious mid-word spaces from the source digitization (e.g. `te nía` → *tenía*, `bené fico` → *benéfico*, `re trato` → *retrato*). These were **deliberately left unfixed** to avoid corrupting the Kichwa side; a Spanish spell-correction pass is future work. - **`review`-flagged rows (~0.4%).** A small number of rows are misaligned in the source (both fields Spanish, or two languages merged into one field). Filter `quality_flag == "ok"` for a clean subset. - **`entry_type` is heuristic** (token-count based), not a linguistic judgment. - **Dialect.** Content is Napo Lowland (Amazonian) Ecuadorian Kichwa (`qvo`). Highland varieties (e.g. Imbabura `qvi`, Chimborazo `qug`) differ lexically and orthographically; do not assume coverage of them. ## Intended uses Machine translation (Kichwa↔Spanish), bilingual lexicon induction, instruction tuning for low-resource Kichwa, and language-preservation / educational tooling. Given the caveats above, human review is recommended before any user-facing deployment. ## Citation Please cite the **original dictionary** and the **contributing Hub datasets**: ```bibtex @misc{diccionario_kichwa_castellano, title = {Diccionario Kichwa--Castellano (Napo / Amazonian Kichwa)}, author = {{UNICEF} and {DINEIB (Direcci\'on Nacional de Educaci\'on Intercultural Biling\"ue, Ecuador)}}, isbn = {978-92-806-4082-3}, note = {Free educational distribution. Digitized copy: FLACSO Andes digital library, doc. 55476}, howpublished = {\url{https://repositorio.flacsoandes.edu.ec/items/da4f1fd3-1393-4cf6-a746-c02e1026b1e4}} } ``` Contributing datasets: `Carlos89apc/TraductorES_Kichwa` (GPL-3.0), `hcurbelo/Kichwa_ES_001` (Apache-2.0), `hcurbelohy/ES_Kichwa` (MIT), all on the Hugging Face Hub. ## Maintenance Cleaning and repackaging only — no new linguistic content was created here. Issues, corrections, and especially fluent-speaker review are welcome via the Community tab.