BeitTigreAI/tigre-data-monolingual-text

Tigre Corpus — Two Files

This release is split into two separate files, because they contain two structurally different kinds of text segmentation. Use them accordingly — do not assume every line across both files represents the same kind of unit.

At a glance:

• tig_corpus_newspaper_sentences.txt — genuine clause/sentence-level segments, derived from real punctuation in the source newspaper text.
• tig_corpus_narrative_chunks.txt — fixed-length 25-token chunks from an embedded narrative/story section of the source corpus that contains no punctuation at all. These are not real sentence boundaries.

File 1: tig_corpus_newspaper_sentences.txt

Metric	Value
Segments (lines)	657,474
Total tokens	7,115,733
Vocabulary size	370,526
Hapax legomena	201,929 (54.5% of vocab)
Segment length — min / median / mean / max	1 / 9 / 10.8 / 30
25th / 75th / 90th percentile length	6 / 14 / 21
Segments ending in a real delimiter (፡/።/!:?)	622,864 (94.7%)

How it was built: the source newspaper text is tokenized and re-joined at genuine punctuation boundaries — ፡, ።, !, ?. (፡ is included deliberately: in Eritrean Tigre usage it functions as a real pause/clause marker, not merely decorative word-spacing, per published research on Tigrinya/Tigre punctuation conventions.) A 30-token cap forces a break if no real delimiter is found within that span, to guard against occasional long unpunctuated stretches; this affects only 5.3% of segments.

Median/mean length (9–11 tokens) is consistent with published machine-translation parallel-corpus norms (commonly ~10–25 tokens/sentence across corpora like TED talks, News Commentary, and healthcare parallel corpora).

File 2: tig_corpus_narrative_chunks.txt

Metric	Value
Segments (lines)	185,055
Total tokens	4,626,357
Vocabulary size	316,566
Hapax legomena	180,098 (56.9% of vocab)
Segment length	fixed at 25 tokens (last segment in the file is a 7-token remainder)
Segments ending in a real delimiter	0 (0%)

How it was built: this text comes from one large, continuous narrative/story section embedded in the source corpus (historical accounts and folk-tale-style prose with dialogue) that contains essentially zero standard punctuation — verified directly: across the entire section, only a single stray punctuation mark occurs. Because there is no reliable delimiter to segment on, this file is split into fixed, non-overlapping 25-token windows. Two independent attempts to find better, content-aware break points using a trained n-gram language model were tested and both failed for a well-understood reason: the model has no real examples of what a sentence ending looks like in this narrative style to learn from. The fixed-window approach is the honest, defensible choice given that constraint — not a shortcut.

Recommended use: fine for LLM pretraining (which already operates on fixed context windows regardless of sentence structure). Use with caution for sentence-level tasks such as translation alignment, since lines here do not correspond to complete grammatical units.

Shared notes

• Both files derive from the same underlying cleaned corpus (boilerplate/duplicate removal, punctuation normalization, gemination-mark correction already applied upstream).
• Source: Eritrean Tigre-language newspaper text (the narrative section may originate from a different embedded source within the same corpus, e.g. a serialized fiction/history column).
• Vocabulary/hapax figures above reflect each file's own tokens; they are not directly comparable to a combined-corpus figure since the two files have different segment-length distributions by construction.