lab-corpora-es

Five small Spanish corpora, written by hand for a set of browser-based machine learning demos at jmwebsoluciones.com/lab. Each one exists to be trained on, in the browser, in well under a minute, by models with a few thousand to a few hundred thousand parameters — a tokenizer, a 2D word2vec, a tiny GPT, a retrieval index. None of them try to represent Spanish as a language; they are deliberately tiny and deliberately narrow.

Dataset Summary

These are the source corpora behind four interactive labs: a byte-pair tokenizer trained live in the browser, a skip-gram word2vec model that animates its own training, a mini-GPT with three model-size presets, and an extractive RAG pipeline. All four demos train or index text on the client side, with no server involved, so the corpora had to be small enough for that to be instant and repeatable, and structured enough that a tiny model has a visible chance of learning something from them.

This dataset collects those corpora in one place, split into five themed configurations plus a combined default config with everything concatenated.

Supported Tasks

• Tokenization / subword modeling (vida-estudiante, refranes, recetas): byte-pair encoding and character-level language modeling on short, repetitive Spanish text.
• Word embeddings (campos-semanticos): skip-gram / word2vec-style training, where the point is that four semantic fields with no shared vocabulary should separate into distinguishable clusters.
• Small causal language modeling (refranes, recetas): training a from-scratch mini-GPT (32-64 embedding dimensions) to complete short Spanish sentences.
• Extractive question answering / retrieval (guia-plantas): chunking a document, embedding the chunks, and answering questions by retrieving and quoting real sentences from the source text.

Languages

Spanish (es) only, in every configuration. The sentences use plain, everyday vocabulary — no regional slang beyond what is common in Madrid, no code-switching, no technical jargon outside a handful of machine-learning terms that show up because the topic itself is machine learning.

Dataset Structure

Data Fields

Every row across every configuration has the same three fields:

• text (string): one sentence, one recipe step, or one paragraph, depending on the configuration (see below).
• source_demo (string): the slug of the lab demo the text was written for — one of mini-gpt, word2vec, rag, bpe.
• theme (string): the semantic or topical grouping inside that demo — one of refranes, recetas, animales, comida, lugares, clima, guia-plantas, vida-estudiante.

Configurations

config	rows	granularity	source_demo	theme(s)
default	237	—	all of the below, concatenated in the order listed	all
refranes	26	one proverb per row	mini-gpt	refranes
recetas	14	one recipe sentence per row	mini-gpt	recetas
campos-semanticos	145	one short sentence per row	word2vec	animales, comida, lugares, clima
guia-plantas	12	one paragraph per row	rag	guia-plantas
vida-estudiante	40	one sentence per row	bpe	vida-estudiante

Data Instances

refranes:

{"text": "Más vale tarde que nunca, aunque lo ideal es no llegar tarde jamás.", "source_demo": "mini-gpt", "theme": "refranes"}

recetas:

{"text": "La tortilla de patatas empieza pelando y cortando las patatas en láminas finas, no demasiado gruesas para que se hagan bien por dentro.", "source_demo": "mini-gpt", "theme": "recetas"}

campos-semanticos:

{"text": "el perro come mucho", "source_demo": "word2vec", "theme": "animales"}
{"text": "el pan sabe rico", "source_demo": "word2vec", "theme": "comida"}

guia-plantas:

{"text": "El riego es donde más plantas he perdido, casi siempre por exceso, no por defecto. La regla que uso ahora es meter el dedo dos nudillos en la tierra: si sale seco, riego; si sale húmedo, espero unos días más.", "source_demo": "rag", "theme": "guia-plantas"}

vida-estudiante:

{"text": "el gradiente descendente parece magia hasta que lo programas tú mismo", "source_demo": "bpe", "theme": "vida-estudiante"}

Loading

from datasets import load_dataset

# everything combined
ds = load_dataset("delcenjo/lab-corpora-es")

# one themed configuration
refranes = load_dataset("delcenjo/lab-corpora-es", "refranes")

Curation Rationale

Every corpus here is small and repetitive on purpose. The demos train in the browser, on the visitor's own device, with no backend — a tokenizer or a mini-GPT has to converge to something visibly interesting in tens of thousands of steps, not millions, and a 2-dimensional word2vec model has to separate four topics with only a few hundred sentences to learn from.

campos-semanticos is the clearest case: it is built from four fields (animals, food, city life, weather) that share almost no vocabulary, not even verbs. That is not an oversight — it is what lets a 2-dimensional embedding, trained live with no PCA and no projection, visibly split into four clusters within a few seconds of animation. A more naturally varied corpus of the same size would not separate as cleanly, which would defeat the point of the demo.

refranes and recetas exist because a from-scratch character-level mini-GPT with 32 to 64 embedding dimensions needs text with strong local patterns — proverbs and recipe steps both repeat sentence shapes and vocabulary far more than natural prose does, which gives a tiny model something learnable inside a training run a visitor will actually sit through.

guia-plantas is the one exception to "repetitive": it needed to be a real, coherent piece of writing, long enough to be split into overlapping chunks and searched, with specific enough claims (watering intervals, symptoms, plant names) that an extractive retrieval demo has clearly right and clearly wrong chunks to retrieve for a given question.

vida-estudiante gives the tokenizer demo a different register again — diary-like sentences about studying and living in Madrid — so a visitor training BPE on it sees merges form around a different vocabulary than the proverbs or the recipes.

Source Data

Initial Data Collection and Curation

All five corpora were written by hand by the dataset's author, specifically for these demos. Nothing here was scraped, downloaded, or adapted from a third-party corpus, book, or website, and no text was generated by a language model. refranes restates well-known, public-domain Spanish proverbs with an added clause about studying or coding, rather than quoting any single author's wording. recetas and vida-estudiante describe ordinary cooking and student life in general terms. guia-plantas is the author's own practical notes on keeping houseplants alive in a small flat, written from memory and not copied from any care guide.

Personal and Sensitive Information

None. No names, locations more specific than "Madrid" as a city in general, dates, or any other information that identifies a real, private individual. The habits described in guia-plantas and vida-estudiante are written in generic first person and are not tied to any identifiable person, place, or event.

Where each corpus lives

• refranes and recetas — selectable corpus presets in the mini-GPT lab: jmwebsoluciones.com/mini-gpt. The refranes text is also the default corpus of the single-block transformer demo at jmwebsoluciones.com/lab/entrena-transformer (kept identical on purpose, so a checkpoint trained in either demo starts from the same text).
• campos-semanticos — the live-training word map at jmwebsoluciones.com/lab/word2vec.
• guia-plantas — the sample document of the extractive RAG demo at jmwebsoluciones.com/lab/rag.
• vida-estudiante — the default training corpus of the byte-pair encoding demo at jmwebsoluciones.com/lab/bpe.

Limitations

This is not a general-purpose Spanish corpus and should not be used to evaluate or represent the Spanish language as a whole:

• It is tiny — 237 rows and under 15,000 characters combined across every configuration. Any model trained on it will overfit quickly; that is expected, since the corresponding demos are built to show exactly that happening in real time, not to produce a useful language model.
• It is heavily biased toward one person's life: a student in Madrid, studying machine learning, who cooks and keeps houseplants. Vocabulary, register, and subject matter all reflect that single perspective. This narrowness is a deliberate feature of the source material (see Curation Rationale above), not an attempt at representativeness.
• campos-semanticos sentences are template-generated combinations of a small set of nouns and verbs within each field, not naturally occurring text — they exist to be learnable, not to be realistic.
• refranes restates common proverbs rather than quoting a single attributable source, so it should not be treated as a proverb-collection reference corpus.
• No validation or test split is provided. These corpora were built for live, visitor-driven training runs in a browser tab, not for benchmarking.

Licensing Information

Released under CC BY 4.0. Attribution: delcenjo.

Citation Information

@misc{lab-corpora-es,
  author = {delcenjo},
  title = {lab-corpora-es: small hand-written Spanish corpora for browser ML demos},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/delcenjo/lab-corpora-es}}
}