algerian-nlp/DziriAlign

DziriAlign

1,000 preference pairs (prompt, chosen, rejected) for aligning language models with Algerian Darja and its sociocultural norms, from the Algerian NLP Collective. Counted 2026-09-17 via the Hub datasets-server (/info?dataset=algerian-nlp/DziriAlign: 1,000 train rows) and re-counted row-by-row with datasets streaming (load_dataset("algerian-nlp/DziriAlign", split="train", streaming=True): 1,000 rows).

The default config answers: when two replies compete, which one sounds like a real Algerian — for market bargaining, daily social life, family norms, and local humor.

from datasets import load_dataset

ds = load_dataset("algerian-nlp/DziriAlign", "default", split="train")

default — preference pairs in four sub-domains

1,000 rows, all in train: 250 market negotiation, 250 socio-linguistic, 250 sociocultural, 250 humor (author sub-domain counts, not recounted by us — no sub-domain field ships in the file).

The gap this config fixes, with the measurement: standard models answer Algerians in formal MSA or culturally flat English. Each row pairs a chosen reply (authentic Darja, local norms) against a rejected reply (formal, robotic, or culturally tone-deaf), ready for DPO, RLHF, or SFT eval. 802 of 1,000 prompts (80.2%, counted by streaming filter len(prompt) > 50) exceed 50 characters — prompts are scenario-length, not single sentences.

No row id field exists: rows join to nothing and have no stable key — key on row order only, which breaks across re-downloads. No sub-domain column exists despite four advertised subsets, so per-subset analysis requires re-deriving the split yourself. What breaks if ignored: any claim of the form "models fail at humor" cannot be measured from this file without your own classifier for which rows are humor.

Scoring protocol

Preference win-rate of a judge (human or model) for chosen over rejected, reported with the judge identity and prompt template. Win-rates under different judges are NOT comparable. Unmeasured: no baseline win-rate has been scored by us on this protocol.

Results

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

Usage

default: Algerian preference alignment (DPO/RLHF/SFT-eval).

from datasets import load_dataset

ds = load_dataset("algerian-nlp/DziriAlign", "default", split="train")
ds[0]
# prompt: Khoya, had sabat chhal mellekher? Ghali chwiya.
# rejected: The listed price is final and non-negotiable as per our store policy. ...
# chosen: Wlh ya khouya ghir rani nbi3 fih b ras malou, ... het 4000 w bsahtek.

# Scenario-length prompts (measured 2026-09-17, streaming, 1,000 rows scanned):
subset = ds.filter(lambda row: len(row["prompt"]) > 50)
len(subset)  # 802

What this dataset does NOT contain — and what breaks if you assume it does. No annotator ids, no agreement scores, no sub-domain labels, and no per-row source id, licence, or retrieval date. Do not assume chosen passed any human review — authorship and review process are undocumented.

Script & code-switching distribution

Unmeasured. No per-script counts were published and none were counted by us. Prompts mix Latin Arabizi and Arabic script by design; rejected replies are disproportionately English and formal MSA, which a preference model can exploit as a length-and-language shortcut rather than learning Darja norms.

Provenance and limits

• —The default config is authored original content (not derived) and inherits MIT from the dataset authors. The sentences are theirs, not ours.
• —Per-row provenance is missing: rows carry no source id, no licence, and no retrieval date.
• —Filtration audit. Unmeasured by us. No annotator count, agreement rate, or review pass is published.
• —Bounded, not cleared. Author-stereotyped "authentic" replies cannot be excluded: one author's idea of Darja norms is presented as ground truth with no community validation measurement.
• —Out of scope. Berber (Tamazight)-matrix utterances and eastern/Saharan regional norms were not targeted; treating chosen as pan-Algerian ground truth would corrupt any generalisation claim.
• —Decontamination of any training corpus against this dataset is unmeasured.

Citation

@misc{algerian_nlp_dzirialign,
  title  = {DziriAlign: 1,000 Algerian Darja preference pairs},
  author = {Touati, Kamel and Algerian NLP Collective},
  year   = {2026},
  url    = {https://huggingface.co/datasets/algerian-nlp/DziriAlign}
}

No DOI or arXiv ID exists yet. No upstream sources to cite: content is authored, not derived.

Licence

MIT, per the dataset authors. You may use, modify, and distribute it provided the source is cited. Note: this repo previously shipped with no YAML frontmatter at all, so the Hub showed no licence or language — this card adds the missing mit frontmatter to match the licence stated in the body.