--- pretty_name: Data-Quality (270m generation probes, early copy) license: other license_name: gemma-3-base-model-outputs-apache-claim-unverified task_categories: [text-classification] size_categories: [1K Hub repository `BeardedMonster/Data-Quality`, **public** as of the 2026-11-15 inventory snapshot, 1.8 MB on `main`, 5 files. No card existed on the Hub before this one. The maintained version is `AletheiaAING/Data-Quality`. An early, small public copy (1.8 MB, 2,888 rows in two folders) of two of the model-generated families of [`AletheiaAING/Data-Quality`](https://huggingface.co/datasets/AletheiaAING/Data-Quality), which holds the full card, the 493,831-row graded-degradation set and the diacritics sets: * `270m_in_lang/` (1,641 rows, 68 languages): continuations written by the base model `google/gemma-3-270m` from real in-language corpus prompts (12-35 words); * `270m_pool_b/` (1,247 rows, 20 languages): English-instruction prompts against the base model; about 99% of outputs are English boilerplate with an African language code, kept as a *floor class* for quality estimators. Row counts are from the maintained card. The file schema (40 columns) is the same as the quality tables: `lang`, `language_name`, `prompt`, `prompt_style`, `n_prompt_words`, `text`, `model`, `provenance`, `script`, `degradation`, `severity`, `is_clean`, `n_words`, `tier1_*` (gates: repetition, is_text, length_ok, not_boilerplate, low_repetition), `tier2_*` (character-run, adjacent-duplicate, orthography, vocabulary, complexity, oversimplification, discourse, adjacency, passive, language consistency, burstiness), readability statistics (`fk_grade`, `flesch_ease`, `words_per_sentence`, `syllables_per_word`, `function_word_ratio`, `content_word_ratio`, `content_ttr`), `llm_judge_prompt`, `prompt_system`, `seed_source`, `source_box`. Example (first row, cut): `lang: swh`, `prompt_style: corpus_continuation`, `model: google/gemma-3-270m`, `provenance: generated_gemma-3-270m`, text a repetitive Swahili continuation. The language list (the maintained card lists 71 ISO codes for all configs; 68 in the in-language config, 20 in the pool config) is not repeated here because the per-config code list is not recorded in this copy: **unknown**. ## How to load ```python from datasets import load_dataset a = load_dataset("BeardedMonster/Data-Quality", "270m_in_lang", split="train") b = load_dataset("BeardedMonster/Data-Quality", "270m_pool_b", split="train") ``` ## How it was used in this project Training and evaluation material for the data-quality detectors (degradation detector AUC 0.78 on farm output; discriminator v2 of the quality-filter pipeline; `translate_pipeline/QF_PIPELINE.md`) and the diacritics model, via the maintained `AletheiaAING/Data-Quality`. ## Known issues, contamination and licence lineage * **Model-generated text** from `google/gemma-3-270m`, a base model that frequently repeats itself and mixes languages; rows are labelled by construction, not by quality review. Gemma Terms of Use apply to the outputs. * The `apache-2.0` licence field is inherited from the first publication and was not verified per source (the prompts are real corpus sentences from mixed upstream sources). * No benchmark screening of the prompt sentences against FLORES+/NTREX/MAFAND is documented. No NLLB-200-3.3B output and no TranslatePsy data is documented in this dataset (checked against `paper_385m/DISCLOSURES.md`, `docs/DATA_INVENTORY.md` and `reports/clean_noncommercial_by_language_2026-10-28.csv`; this is a statement about the project documents, not a scan of the rows). ## Visibility and maintenance Public, 1.8 MB. Superseded by `AletheiaAING/Data-Quality`; consider deleting this copy to avoid two public versions with different licence claims. ## Citation ```bibtex @misc{sabiyarn_data_quality, title = {Data-Quality}, author = {TODO(owner): author list (Aletheia AI)}, year = {2026}, note = {Citation placeholder; paper in preparation. Hub repository BeardedMonster/Data-Quality} } ```