BeitTigreAI/tigre-data-monolingual-text

Tigre Corpus

Monolingual text in Tigre (ትግረ; ISO 639-3 tig; Ethiopic script, tig_Ethi), a Semitic language of Eritrea and Sudan.

The corpus has two training files with different kinds of segmentation, plus one held-out test file. Do not assume every line is the same kind of unit.

At a glance

• tig_newspaper_cleaned.txt: real sentence/clause segments, split at punctuation in newspaper text.
• tig_narrative_cleaned.txt: fixed 25-token chunks from a narrative section that has no punctuation. These are not sentence boundaries.
• tigre_test.txt: 2,000 held-out lines for evaluation. Do not train on this file.

Contents

File	Split	Lines	Tokens	Characters
tig_newspaper_cleaned.txt	train	614,682	7,002,396	28.2M
tig_narrative_cleaned.txt	train	181,070	4,526,732	19.4M
Train total		795,752	11,529,128	47.6M
tigre_test.txt	test	2,000	28,285	0.1M

Tokens are whitespace-separated, and punctuation marks are separate tokens. All files are plain UTF-8, Unicode NFC, one segment per line.

Load with the 🤗 datasets library:

from datasets import load_dataset
repo = "BeitTigreAI/tigre-data-monolingual-text"

ds = load_dataset(repo)                       # train (both files) + test
news = load_dataset(repo, "newspaper")        # newspaper sentences only
narrative = load_dataset(repo, "narrative")   # narrative chunks only

print(ds)
print(ds["train"][0]["text"])

Test split (held out)

tigre_test.txt holds 2,000 lines drawn at random from the cleaned corpus: 1,600 newspaper lines (80%) and 400 narrative lines (20%). The sample is weighted towards newspaper sentences because short text is where language identification is hardest. These lines are removed from both training files, so the splits do not overlap.

Please keep this file out of training. It lets anyone, including language-identification projects such as GlotLID, check whether a model trained with this corpus has really improved on Tigre.

File 1: tig_newspaper_cleaned.txt

Metric	Value
Segments (lines)	614,682
Total tokens	7,002,396
Vocabulary size	398,372
Hapax legomena	230,076 (57.8% of vocabulary)
Segment length: min / median / mean / max	1 / 10 / 11.4 / 30 tokens
25th / 75th / 90th percentile length	6 / 14 / 22 tokens
Segments ending in a real delimiter (፡ ። ! ?)	579,732 (94.3%)

How it was built. The newspaper text was tokenized and re-joined at genuine punctuation boundaries: ፡, ።, ! and ?. ፡ is included deliberately, because in Eritrean Tigre usage it works as a real pause or clause marker, not only as decorative word spacing. If no delimiter appears within 30 tokens, a break is forced, which guards against occasional long unpunctuated stretches.

The median length of about 10 tokens is in line with common sentence-level corpora.

File 2: tig_narrative_cleaned.txt

Metric	Value
Segments (lines)	181,070
Total tokens	4,526,732
Vocabulary size	300,665
Hapax legomena	166,923 (55.5% of vocabulary)
Segment length	25 tokens (one 7-token remainder)
Segments ending in a real delimiter	0 (0%)

How it was built. This text comes from one large, continuous narrative section of the source material, with historical accounts and folk-tale-style prose including dialogue. The section has essentially no standard punctuation; only a single stray mark occurs in the whole section. With no reliable delimiter to segment on, the file is split into fixed, non-overlapping 25-token windows. Two attempts to find content-aware break points with a trained n-gram language model failed, because the model had no examples of sentence endings in this narrative style to learn from. Fixed windows are therefore the honest choice here.

Recommended use. Suitable for language-model pretraining and language identification. Use with caution for sentence-level tasks such as translation alignment, because lines do not correspond to complete grammatical units.

Collection and cleaning

1. Upstream: boilerplate and duplicate removal, punctuation normalisation and gemination-mark correction.
2. Segmentation: as described for each file above.
3. Release cleaning:
   • Unicode NFC normalisation.
   • Lines that were broken text from PDF or font extraction were removed (6,971 lines). These included Latin mojibake, digits fused into words, and runs of isolated syllables.
   • Exact duplicates were removed, compared after normalisation (8,071 lines).
   • Near-empty lines under 5 characters were removed (11,410 lines).
   • The 2,000-line test split was set aside.

Headlines, bylines, dates and page furniture such as mastheads are genuine Tigre text and were kept.

Source and licence

• Source: Eritrean Tigre-language newspaper material donated by the Ministry of Information of Eritrea for open use, as part of an initiative to help languages of Eritrea thrive in digital technology. The narrative section is an embedded part of the same material, for example a serialized fiction or history column.
• Licence: Creative Commons Attribution 4.0 International (CC BY 4.0). The donor provided the text for open use without restrictions.
• Attribution: please credit "Tigre Corpus, text donated by the Ministry of Information of Eritrea" and cite this dataset (below).

Intended uses

• Language identification, for example improving Tigre in GlotLID.
• Language modelling and pretraining.
• Research on Tigre and related Ethio-Semitic languages.

Limitations. The text is mostly news and one narrative collection, so other registers such as social media, conversation and religious text are under-represented. Short newspaper lines include headlines, names and sports results.

Notes on the statistics

Vocabulary and hapax figures are computed per file on whitespace tokens. They are not directly comparable across files, because the two files have different segment-length distributions by construction.

Citation and contact

@misc{beittigreai_tigre_monolingual_2026,
  title        = {Tigre Monolingual Text Corpus},
  author       = {{BeitTigreAI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/BeitTigreAI/tigre-data-monolingual-text}},
  note         = {Text donated by the Ministry of Information of Eritrea. CC BY 4.0}
}