DeepLatent bilingual tokenizer-data

Balanced English/Arabic corpus for tokenizer training. The two languages carry essentially the same number of Unicode codepoints, so a BPE/WordPiece tokenizer trained on this corpus sees equal representation of both languages by character content.

Composition

Slice	Source	Filter	Rows	Chars
English	almaghrabima/deeplatent-hq-merged-dedup-token-counts	GlotLID language == "English"	3,980,035	12,377,830,142
Arabic	AdaMLLab/AraMix-HQ	mmbert_score >= 0.2784 ∧ language != "English"	2,380,570	12,380,271,596
Total			6,360,605	24,758,101,738

The Arabic threshold mmbert_score >= 0.2784 was chosen so the Arabic char count matches the English char count (balancing the two languages). This yields the top ~7% highest-scoring Arabic content in AraMix-HQ.

Language labels come from GlotLID (cis-lmu/glotlid) run on the first 2000 characters of each document. The HQ corpus is fully labeled; the AraMix-HQ source was partially labeled (~38.5% of shards) — remaining Arabic rows in this merged release default to "Arabic" since labeled AraMix-HQ shards were 96.4% Arabic.

Schema

Column	Type	Description
text	string	Document text
source	string	Original sub-source (e.g. ar, en, lightonai/ArabicWeb24)
language	string	GlotLID label: "English", "Arabic", or raw lang_Script
origin	string	"hq" or "aramix"

File layout

• en_00000.parquet … en_00222.parquet — 223 English shards
• ar_00000.parquet … ar_00178.parquet — 179 Arabic shards

Each file is zstd-compressed parquet.