algerian-nlp/TinyStories-Algerian-Darija

TinyStories Algerian Darija

Synthetic short stories in Algerian Darja paired with their English originals, for Darija language modeling and translation, from the Algerian NLP Collective. The Hub datasets-server reports 11,326 train rows (/info?dataset=algerian-nlp/TinyStories-Algerian-Darija, 2026-09-17), independently confirmed by the build ledger processed_story_ids.json in this repo: 11,326 unique story ids (0 to 11,492, non-contiguous).

The default config answers: what does simple narrative Darija look like at scale when generated from English TinyStories-style prompts.

from datasets import load_dataset

ds = load_dataset("algerian-nlp/TinyStories-Algerian-Darija", "default", split="train")

default — synthetic stories across 31 parquet files

11,326 rows per the datasets-server count, all in train, across 381 files (data/train_batch_000001.parquet through data/train_batch_000381.parquet, file list from the Hub API 2026-09-17). No dev or test split ships with this repo.

The gap this config fixes is stated, not measured by us: simple narrative Darija for small-model training. Row 0 carries darija_word_count 55 against english_word_count 134 with generator gemini-3.5-flash and timestamp 2026-09-01T18:33:34Z — generator identity and date are per-row fields, so generator mix is measurable per file once the schema defect below is repaired.

Schema defect, binding. Late files (observed at train_batch_000037.parquet in streaming storage order) carry moral_lesson_darija, theme, age_group INSTEAD of model and timestamp. Column names do not match across files, so a full load_dataset scan fails with a CastError (reproduced 2026-09-17). Consequence: until the owner re-exports one schema, load per-file with pyarrow or pin the early-file subset, and treat the 11,326 server count as the file-sum, not as a loadable row count. What breaks if ignored: any ds.filter or full iteration over the repo as one dataset raises instead of returning rows.

Scoring protocol

Language-modeling perplexity or translation BLEU/chrF on a split you define, with tokenizer, signature, and seed beside the number. Unmeasured: no baseline scores have been produced by us on this protocol.

Results

Unmeasured. No baseline results scored by us in the protocol above.

Usage

default: synthetic narrative Darija for modeling and translation.

from datasets import load_dataset

ds = load_dataset("algerian-nlp/TinyStories-Algerian-Darija", "default", split="train")
ds[0]
# id: 0
# story_algerian_darija: واحد النهار، كانت كاين واحد البنيتة شاطرة ...
# story_english_original: One day, a little girl named Lily found a needle ...
# darija_word_count: 55 — english_word_count: 134
# model: gemini-3.5-flash — timestamp: 2026-09-01T18:33:34Z

Filtered views over the whole repo are unmeasured: the schema defect above aborts full scans. Per-file filtered counts are possible and left to the reader until the schema is unified.

What this dataset does NOT contain — and what breaks if you assume it does. This repo previously shipped with NO card at all. There is no dev or test split, no human review flag, no uniform schema, and no per-row source id, licence, or retrieval date. Do not assume all rows share row 0's generator — model varies per row where present and is absent where the late schema applies.

Script & code-switching distribution

Unmeasured. No per-script counts were published and none were counted by us; the observed Darija side is Arabic-script.

Provenance and limits

• —The default config is synthetic output of English-prompted generators (observed: gemini-3.5-flash), not human Algerian writing. Human provenance is absent by construction.
• —Per-row provenance is missing: rows carry no source id, no licence, and no retrieval date; timestamp is generation time, not retrieval.
• —Filtration audit. Unmeasured by us and unstated by the owner. No review pass, no human rating, and no de-duplication report exist.
• —Bounded, not cleared. Generator hallucinations, MSA-flavoured phrasing, and English calques cannot be excluded: no human validation measurement ships with the set.
• —Out of scope. Real spoken Darja, Arabizi, and code-switching were not targeted; this set measures synthetic narrative style only.
• —Decontamination of any training corpus against this dataset is unmeasured.

Citation

@misc{algerian_nlp_tinystories_algerian_darija,
  title  = {TinyStories Algerian Darija: synthetic Darija stories with English originals},
  author = {Algerian NLP Collective},
  year   = {2026},
  url    = {https://huggingface.co/datasets/algerian-nlp/TinyStories-Algerian-Darija}
}

No DOI, arXiv ID, or named author is published on the repo. Upstream: the English TinyStories family of synthetic story prompts — cite the generator family alongside this set.

Licence

Unknown. No licence tag or licence statement ships with this repo, so no grant can be quoted here. Do not redistribute derivatives until the owner declares a licence; this card sets unknown rather than inventing one.