Ha-Te-21/tigrinya-corpus

Tigrinya Training Corpus

A large, multi-source Tigrinya-language text corpus assembled for continued pretraining (CPT) and instruction-tuning of LLMs on Tigrinya. It combines natively-authored Tigrinya text (news, religious texts, literature, legal documents, social media, OCR'd books), machine-translated text (English → Tigrinya via NLLB, quality-filtered), and synthetic instruction/QA data.

⚠️ Draft card. Sizes and subset descriptions below were generated from the local file layout. License, exact train/held-out splits, and a few subset provenance details (marked below) need to be confirmed/filled in by the dataset author before publishing.

Dataset Summary

• Language: Tigrinya (ti), Ethiopic script
• Total size: ~2.6 GB across 32 source subsets, plus a ~1.2 GB merged file of the translated subsets (see Note on all_translated_pretrain.jsonl — it overlaps with 5 of the subsets below, not additional data)
• Format: JSON Lines (.jsonl), UTF-8
• Primary use case: continued pretraining / domain adaptation of a base LLM on Tigrinya, with a smaller instruction-tuning slice (QA, exam-style multiple choice)

Dataset Structure

Two record schemas appear across the corpus:

1. Plain-text pretraining records (most subsets):

{"text": "...", "source": "...", "date": "2022-11-08"}

Some subsets add extra fields, e.g. the translated subsets include domain, word_count, quality_score, ethiopic_ratio, amharic_rate, n_chunks, block_id; the MMLU-Pro subset adds category, question_id, answer, answer_index, roundtrip_question_score, and the original English question.

2. Chat/instruction records (tigrigna_qa):

{"messages": [
  {"role": "system", "content": "<lang:tigrigna>\nYou are a helpful assistant. Respond in Tigrigna."},
  {"role": "user", "content": "..."},
  {"role": "assistant", "content": "..."}
]}

Data Fields (plain-text schema)

Field	Type	Description
text	string	The Tigrinya text sample
source	string	Origin identifier (site, corpus name, or synthetic:* / *_translated tag)
date	string	Publication date, where available (news/scraped subsets)
domain	string	Topical domain, translated subsets only
quality_score	float	Automated quality/filter score, translated subsets only
ethiopic_ratio	float	Fraction of characters in Ethiopic script, translated subsets only

Subsets

Subset (folder)	Size	Type	Description
all_translated_pretrain.jsonl (root)	1.21 GB	merged	Deduplicated merge of the 5 *_translated subsets below — use either this or the individual folders, not both
BBC_Tigrinya	53.95 MB	native	Scraped BBC Tigrinya news articles
BBC_s2	0.97 MB	native	Additional BBC-sourced batch
BBC_Amharic_synthetic	0.07 MB	synthetic	Synthetic/derived from Amharic BBC content
Bible_Scripture_synthetic	2.33 MB	synthetic	Scripture text, synthetic/derived
Chilot_Laws	9.10 MB	native	Ethiopian legal/law text (Chilot)
docx_s2	0.34 MB	native	Text extracted from Word documents
Educational_Books_synthetic	32.69 MB	synthetic	Synthetic educational book content (includes filtered + rejected variants alongside the training-ready file)
english_news_translated	14.29 MB	translated	English news, NLLB-translated to Tigrinya, quality-filtered
ERIPM	2.29 MB	native	ERIPM corpus
glocr	5.99 MB	OCR	OCR'd text, unsegmented
GoodTigrignaBooks_s2	13.75 MB	native	Curated Tigrinya book text
gutenberg_tigrinya	728.82 MB	unconfirmed	Two files (gutenberg_tigrinya.jsonl + a _FILTERED variant). Provenance needs confirming — this is a separate/newer pass from gutenberg_translated below (per the pipeline report, a newer parallel_en_ti.tsv run wasn't included in the translation-quality report)
gutenberg_translated	464.70 MB	translated	Project Gutenberg text, NLLB-translated, quality-filtered (88.8% pairs accepted — see pipeline stats)
Haddas_Ertra_OCR	39.76 MB	OCR	OCR'd from Haddas Ertra (Eritrean newspaper)
hplt	326.45 MB	native	Web-crawled Tigrinya from the HPLT corpus
InternetArchive_s2	1.00 MB	native	Text sourced from Internet Archive
masakhanews	7.63 MB	native	MasakhaNEWS Tigrinya split
MMLU_Pro_synthetic	8.87 MB	synthetic	MMLU-Pro exam questions translated/adapted to Tigrinya (multiple-choice, with English originals retained)
news_tigrinya	25.09 MB	native	General Tigrinya news corpus
openstax_translated	43.38 MB	translated	OpenStax textbook content, NLLB-translated (84.4% accepted)
open_textbooks_translated	673.96 MB	translated	Open textbook content, NLLB-translated (88.1% accepted) — largest translated subset
Orthodox_left_s2 / Orthodox_s2	3.34 MB / 3.81 MB	native	Ethiopian/Eritrean Orthodox religious text
telegram_s2	17.80 MB	native	Scraped Telegram channel text. Known issue: contains word-merging artifacts and U+FFFD encoding corruption in places — recommend re-cleaning before training use
TGHAT	1.66 MB	native	TGHAT corpus
TigrayGenocide_synthetic	13.82 MB	synthetic	Synthetic content documenting the Tigray genocide — see Sensitive Content below
tigrigna_qa	7.56 MB	instruction	Chat-format QA pairs, system/user/assistant roles
various_left_s2 / various_s2	11.61 MB / 1.84 MB	native	Mixed/uncategorized native text batches
wikipedia_native	0.72 MB	native	Native Tigrinya Wikipedia articles
Wikipedia_s2	0.76 MB	native	Additional Wikipedia-sourced batch
wikipedia_translated	42.53 MB	translated	English Wikipedia, NLLB-translated (90.5% accepted)
wurayna_s2	1.80 MB	native	Wurayna-sourced batch
YouTube_EthioForum_synthetic	15.32 MB	synthetic	Synthetic content derived from YouTube/EthioForum discussions

"s2" batches are the dataset author's internal naming for a second collection pass; exact source detail can be added here if useful for users.

Translation Pipeline & Quality

The *_translated subsets were produced via NLLB machine translation (EN→TI) followed by an automated quality filter (no LaBSE; filtering via entity-echo, sentence-ratio, and length-ratio heuristics). Aggregate stats across all 5 translated sources:

• 3,240,077 lines read → 2,851,317 pairs accepted (88.0%)
• 380,315 final deduplicated blocks in the combined file
• Rejection reasons (by volume): length_ratio (2.64%), ethiopic_ratio (3.44%), english_passthrough (1.39%), entity_echo (1.38%), sentence_ratio (1.06%), toc_line (1.48%), duplicate_exact (0.54%), remainder minor categories

Per-source acceptance rates: English news 64.1%, Gutenberg 88.8%, open textbooks 88.1%, OpenStax 84.4%, Wikipedia 90.5%.

Note on all_translated_pretrain.jsonl

The root-level all_translated_pretrain.jsonl (1.21 GB) is the deduplicated merge of english_news_translated + gutenberg_translated + openstax_translated + open_textbooks_translated + wikipedia_translated. Loading both the merged file and the individual subsets will duplicate ~1.2 GB of data — pick one path depending on whether you need per-source domain/quality_score metadata (individual folders) or a single ready-to-stream file (merged).

Sensitive Content

The TigrayGenocide_synthetic subset contains content documenting the 2020–2022 Tigray war and associated atrocities. This is included for historical/documentary completeness in a corpus aimed at Tigrinya-speaking communities, not as generic training filler. Users building general-purpose assistants may want to weight, filter, or add safety handling around this subset depending on their downstream use case.

Known Data Quality Notes

• telegram_s2: word-merging and occasional U+FFFD replacement-character corruption present in the raw text; recommend byte-level verification and re-cleaning before use in a final training mix.
• gutenberg_tigrinya provenance vs. gutenberg_translated needs a one-line clarification from the author (see table above).
• Rejected/filtered companion files (*_rejected.jsonl, *_FILTERED.jsonl, *_rejected_summary.txt) are kept alongside several subsets for auditability — the *_training_ready.jsonl / *_pretrain.jsonl file in each folder is the one intended for actual training.

Licensing

(To fill in.) Source material spans BBC, Wikipedia, Project Gutenberg, OpenStax, Orthodox Church texts, Telegram channels, OCR'd newspapers, and synthetic generations — likely a mix of licenses. Recommend stating research-use-only or per-source licensing until each subset's terms are confirmed.

Citation

(Add citation / author attribution here.)