BeardedMonster/sabiyarn-3B

SabiYarn 3B tokenizer

Byte-level BPE tokenizer with a 131,072-token vocabulary (2^17) for English, French, Arabic and African languages. This repository holds the tokenizer only; no SabiYarn 3B model weights exist yet.

It is the successor of the 65,536-token tokenizer of SabiYarn 0.5B, trained on a language-balanced sample of the project's pretraining corpus.

Properties

• Vocabulary 131,072 = 180 special tokens + 256 byte symbols + 130,636 merges. Because ids go above 65,535, tokenized corpora need uint32 (4 bytes per token), not uint16.
• Special tokens are unchanged. The 180 special tokens (language, task, chat-role, reasoning, tool-call and audio tokens, and the quoted JSON keys) keep exactly the ids 0-179 they have in SabiYarn 0.5B, with the same flags, so data, chat template and tool tokens written for the 0.5B tokenizer carry over. The chat template (chat_template.jinja) is identical.
• NFC normalizer. Decomposed (NFD) and composed (NFC) spellings of a word get identical ids. Text that is already NFC round-trips exactly (decode(encode(text)) == text); other input decodes to its NFC form.
• No OOV. All 256 byte symbols are in the vocabulary.
• Checked on 20,910 evaluation texts (73 languages): NFC text round-tripped exactly in 20,910 of 20,910, and NFD versions decoded to the NFC string in 20,910 of 20,910.

How it was trained

• Corpus. 4.91 GB of text (14.35 million documents, 71 languages, 1.57 billion tokens under the 0.5B tokenizer) sampled at random from the project's deduplicated, language-verified pretraining corpus.
• Balance. English 8%, French 4% and Arabic 3% of the budget; the remaining 85% allocated UniMax-style with one epoch per language (languages with little data taken whole, the larger ones capped at roughly 25-30 million tokens each, about 55-115 MB of text).
• Decontamination. Documents containing a FLORES+ devtest sentence, or equal to a FineWeb-2 evaluation document used below, were removed from the training sample (50,649 documents).
• Training. Hugging Face tokenizers BpeTrainer, min_frequency 2, NFC normalizer, the same pre-tokenizer as SabiYarn 0.5B (whitespace and punctuation split followed by byte-level BPE), the 180 special tokens pinned to ids 0-179. Training took 17 minutes.
• Effect on the pretraining mix. On fresh held-out text sampled from the same corpus, weighted by the language mix of the 40-billion-token training bin, text needs 14.8% fewer tokens than with the 0.5B tokenizer (English 4% fewer, French 30%, Zulu 17%, Wolof 19%, Kikuyu 38%, Kabiye 45%; Amharic unchanged).

Tokenizer comparison across supported languages

Tokens per word for this tokenizer against other tokenizers on the same text, lower is better. Measured 2026-10-20.

How it was measured

• Metric. Tokens per whitespace-separated word, every text NFC-normalised first and no special tokens added. Premium vs English is a language's token count divided by the English count for the same parallel sentences (FLORES+ only). Unknown tokens is the share of output tokens equal to the tokenizer's UNK id; tokenizers with a high share lose characters, which lowers their counts.
• Languages. The 86 language tags in the special tokens plus Arabic. 73 had accessible text and are evaluated; no usable text was found for ijo, ful, ibb, anw, orx, igl, snq, nup, kln, lch, son, kon, vai, mey. Core languages (n=25) are the languages named on the SabiYarn 0.5B card; Other tags are the rest.
• Text. F = FLORES+ devtest (1,012 parallel sentences per language; 51 languages including English and Arabic). W = FineWeb-2 web text for languages FLORES+ lacks (22 languages; 25 to 600 documents of up to 2,000 characters each).
• Reused values. The values for every tokenizer except this one come from the SabiYarn 0.5B model card measurement of 2026-10-17 and were not re-measured. As a control, re-running the 0.5B tokenizer on the regenerated texts reproduced its card values to within 0.005 on all 73 languages. The "Best on # langs" and "Within 10% of best" columns were recomputed with this tokenizer included.
• Caveats. Whitespace words are not comparable across languages, so compare tokenizers within a row. The W texts come from the same family of sources as part of the training corpus, so this tokenizer's advantage there may be somewhat inflated; on the held-out F languages alone the mean is 1.70 (this tokenizer), 1.73 (Serengeti), 1.82 (NLLB-200), 2.07 (SabiYarn 0.5B), 2.37 (Gemma 3).

Summary (mean tokens per word)

Tokenizer	Vocab	Core langs (n=25)	Other tags (n=48)	All (n=73)	Premium vs English (FLORES+, n=50)	Best on # langs	Within 10% of best	Unknown tokens	English
SabiYarn 3B (131k)	131,072	1.61	1.76	1.71	1.39	42	61	0.0%	1.27
Serengeti	250,004	1.63	1.79	1.74	1.41	22	56	2.6%	1.28
SabiYarn 0.5B	65,536	1.78	2.46	2.23	1.66	0	8	0.0%	1.31
AfriTeVa v2	150,100	1.83	2.17	2.05	1.63	2	20	5.3%	1.31
NLLB-200	256,204	1.83	1.89	1.87	1.35	2	33	0.8%	1.40
AfriBERTa	70,005	2.22	2.52	2.42	1.54	1	6	0.0%	1.64
AfroXLMR	250,002	2.29	2.33	2.31	1.73	0	2	1.5%	1.40
Gemma 3	262,145	2.37	2.42	2.40	1.99	0	3	0.0%	1.24
Gemma 4	262,144	2.37	2.42	2.40	1.99	0	3	0.0%	1.24
Llama 4	201,135	2.44	2.56	2.52	2.06	0	3	0.0%	1.24
Morena 1.5B	65,536	2.53	2.60	2.57	1.84	2	8	0.0%	1.41
BLOOM	250,680	2.55	2.42	2.47	1.99	1	5	0.0%	1.25
GPT-4o / GPT-5 (o200k)	200,019	2.61	2.49	2.53	2.07	2	4	0.0%	1.23
gpt-oss (o200k_harmony)	201,088	2.61	2.49	2.53	2.07	2	4	0.0%	1.23
Qwen3	151,669	2.85	2.64	2.71	2.23	0	1	0.0%	1.26
Aya Expanse	255,029	2.96	2.64	2.75	2.27	1	4	0.0%	1.23
DeepSeek-V3	128,815	3.06	2.78	2.87	2.35	1	2	0.0%	1.23
InkubaLM 0.4B*	61,788	3.18	2.58	2.79	2.04	0	0	0.0%	1.40
Mistral Nemo	131,072	3.25	2.85	2.99	2.38	0	1	0.0%	1.27
Llama 3.2	128,256	3.28	2.89	3.02	2.48	0	2	0.0%	1.24
GPT-4 (cl100k)	100,277	3.41	2.95	3.11	2.55	0	1	0.0%	1.24
Llama 2	32,000	3.70	3.00	3.24	2.31	0	0	0.0%	1.42

Tokenizers and sources

Tokenizer	Vocab	Loaded from
SabiYarn 3B (131k)	131,072	BeardedMonster/sabiyarn-3B
SabiYarn 0.5B	65,536	AletheiaAING/sabiyarn-0.5B
Morena 1.5B	65,536	vamboai/morena-1.5b-instruct
InkubaLM 0.4B*	61,788	OxxoCodes/InkubaLM-Instruct-test (copy of lelapa/InkubaLM-0.4B tokenizer)
Gemma 3	262,145	unsloth/gemma-3-1b-it
Gemma 4	262,144	google/gemma-4-E4B-it
GPT-4 (cl100k)	100,277	tiktoken cl100k_base
GPT-4o / GPT-5 (o200k)	200,019	tiktoken o200k_base
gpt-oss (o200k_harmony)	201,088	tiktoken o200k_harmony
Llama 2	32,000	NousResearch/Llama-2-7b-hf
Llama 3.2	128,256	unsloth/Llama-3.2-1B
Llama 4	201,135	unsloth/Llama-4-Scout-17B-16E-Instruct (mirror)
Qwen3	151,669	Qwen/Qwen3-0.6B
AfriBERTa	70,005	castorini/afriberta_base
AfroXLMR	250,002	Davlan/afro-xlmr-large
Serengeti	250,004	UBC-NLP/serengeti
NLLB-200	256,204	facebook/nllb-200-distilled-600M
AfriTeVa v2	150,100	castorini/afriteva_v2_base
BLOOM	250,680	bigscience/bloom-560m
Aya Expanse	255,029	CohereLabs/aya-expanse-8b
Mistral Nemo	131,072	unsloth/Mistral-Nemo-Instruct-2407
DeepSeek-V3	128,815	deepseek-ai/DeepSeek-V3

* InkubaLM's official repository is gated, so its tokenizer was loaded from a public fine-tune copy (OxxoCodes/InkubaLM-Instruct-test); its vocabulary size (61,788) matches the published figure. Llama 4 comes from a mirror because the official repository is gated. GPT-5 uses o200k_base through tiktoken (encoding_for_model("gpt-5")), and gpt-oss uses o200k_harmony. Gemma 3 and Gemma 4 give identical counts in this test.

Reading the results. On the 25 core languages the 131k tokenizer averages 1.61 tokens per word against 1.78 for the 0.5B tokenizer, and across all 73 languages 1.71 against 2.23. It also edges the specialist African tokenizer Serengeti (250k vocabulary) on all three means (1.61, 1.76 and 1.71 against 1.63, 1.79 and 1.74) and has the lowest count on 42 of 73 languages. It is lower than the 0.5B tokenizer on 72 of 73 languages. The biggest changes are French (2.08 to 1.42, now level with GPT-4o and Gemma 3), the Southern Bantu languages (Zulu 2.48 to 2.09, Xhosa 2.46 to 2.09, Shona 2.27 to 1.98), Kikuyu (2.93 to 1.75), Kabiye (3.28 to 1.83) and Tifinagh (tzm, 12.2 to 2.5). Known gaps: Amharic is worse than the 0.5B tokenizer on FLORES+ (1.91 against 1.78) and Arabic is unchanged; Serengeti and NLLB-200 remain lower on Swahili (1.50 against 1.26), Afrikaans, Somali, Malagasy, Tifinagh and Amharic.

Usage

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("BeardedMonster/sabiyarn-3B")
tok.apply_chat_template(
    [{"role": "user", "content": "Bawo ni?"}], tokenize=False
)

Per-language tokenizer comparison

Tokens per word for each tokenizer on the same text, lower is better; the best value in each row is in bold. Text is the source: F = FLORES+ devtest (1,012 parallel sentences), W = FineWeb-2 web text. Core languages (those named at the top of this card) come first, then the remaining language tags, each alphabetically. Method, caveats and sources are in the sections above.

Language	Code	Text	SabiYarn 3B (131k)	SabiYarn 0.5B	Morena 1.5B	InkubaLM 0.4B*	Gemma 3	Gemma 4	GPT-4 (cl100k)	GPT-4o / GPT-5 (o200k)	gpt-oss (o200k_harmony)	Llama 2	Llama 3.2	Llama 4	Qwen3	AfriBERTa	AfroXLMR	Serengeti	NLLB-200	AfriTeVa v2	BLOOM	Aya Expanse	Mistral Nemo	DeepSeek-V3
Afrikaans	afr	F	1.56	1.70	1.57	2.02	1.79	1.79	1.99	1.66	1.66	2.09	1.98	1.75	2.01	2.45	1.59	1.35	1.61	1.40	2.00	1.75	1.83	1.91
Amharic	amh	F	1.91	1.78	7.49	12.98	3.03	3.03	11.91	8.96	8.96	13.03	11.89	3.73	6.45	1.69	2.33	1.58	2.32	1.75	7.90	9.83	11.93	9.32
Arabic	ara	F	1.89	1.94	4.57	4.53	2.03	2.03	4.19	1.89	1.89	5.41	2.29	2.19	2.29	3.07	1.83	1.59	2.17	2.37	1.60	2.03	1.76	2.23
Chichewa	nya	F	1.67	1.90	2.35	2.54	2.57	2.57	2.88	2.21	2.21	3.18	2.84	2.52	2.90	2.69	2.48	1.90	1.79	1.68	2.47	2.67	2.81	2.85
Efik	efi	W	1.46	1.80	2.44	2.42	2.54	2.54	2.97	2.47	2.47	3.21	2.86	2.61	2.92	2.44	2.75	1.77	2.19	2.10	2.41	2.72	2.96	2.84
English	eng	F	1.27	1.31	1.41	1.40	1.24	1.24	1.24	1.23	1.23	1.42	1.24	1.24	1.26	1.64	1.40	1.28	1.40	1.31	1.25	1.23	1.27	1.23
Ewe	ewe	F	1.48	1.59	2.64	2.57	2.37	2.37	2.98	2.52	2.52	2.83	2.86	2.49	2.59	2.57	2.46	1.64	1.54	2.40	2.31	2.59	2.91	2.69
Fon	fon	F	1.44	1.63	2.70	2.58	2.28	2.28	3.19	2.48	2.48	2.82	3.14	2.48	2.58	2.54	2.51	1.53	1.59	2.40	1.89	2.54	3.10	2.72
Hausa	hau	F	1.32	1.35	1.39	1.57	1.86	1.86	2.15	1.66	1.66	2.33	2.13	1.86	2.13	1.37	1.69	1.28	1.43	1.31	1.92	1.98	2.10	2.09
Igbo	ibo	F	1.41	1.43	1.57	2.48	2.36	2.36	2.69	1.79	1.79	2.93	2.62	2.22	2.64	1.57	2.64	1.53	1.62	1.56	1.89	2.60	2.67	2.63
Kinyarwanda	kin	F	1.70	1.95	1.86	2.60	2.65	2.65	2.88	2.29	2.29	3.16	2.86	2.67	2.90	1.95	2.61	1.82	1.97	1.86	2.15	2.72	2.81	2.86
Nigerian Pidgin	pcm	W	1.40	1.45	1.51	1.56	1.44	1.44	1.52	1.42	1.42	1.69	1.50	1.45	1.53	1.58	1.53	1.38	1.45	1.51	1.45	1.46	1.52	1.51
Oromo	orm	F	1.97	2.16	3.11	3.11	2.95	2.95	3.22	2.58	2.58	3.53	3.20	2.98	3.24	1.89	2.79	1.92	2.22	2.20	3.06	3.01	3.15	3.21
Setswana	tsn	F	1.39	1.56	1.45	1.89	1.84	1.84	2.00	1.66	1.66	2.12	1.99	1.83	2.01	2.03	1.82	1.38	1.43	1.47	1.78	1.90	1.96	1.98
Shona	sna	F	1.98	2.27	1.94	3.03	2.93	2.93	3.34	2.63	2.63	3.59	3.25	2.92	3.36	3.22	2.88	2.17	2.15	1.95	2.85	3.04	3.18	3.25
Somali	som	F	1.66	1.82	2.48	2.45	2.25	2.25	2.57	1.99	1.99	2.87	2.55	2.26	2.59	1.54	1.84	1.45	1.83	1.42	2.40	2.41	2.57	2.58
Southern Sotho	sot	F	1.40	1.55	1.69	1.95	1.89	1.89	2.09	1.71	1.71	2.24	2.07	1.88	2.10	2.13	1.89	1.43	1.47	1.45	1.86	1.98	2.06	2.06
Swahili	swh	F	1.50	1.71	1.54	1.69	2.09	2.09	2.49	1.88	1.88	2.71	2.46	2.12	2.51	1.49	1.67	1.26	1.62	1.40	1.60	2.30	2.37	2.46
Tigrinya	ti	F	1.92	1.97	6.91	11.94	3.46	3.46	10.91	8.27	8.27	11.98	10.90	4.19	5.92	1.91	3.10	1.86	2.27	2.11	7.25	9.01	10.92	8.54
Twi (Akuapem)	twi	F	1.39	1.51	2.20	2.16	2.02	2.02	2.61	1.91	1.91	2.40	2.60	2.14	2.27	2.31	2.21	1.51	1.47	2.09	1.90	2.20	2.55	2.25
Urhobo	urh	W	1.36	1.75	2.17	2.12	2.40	2.40	2.88	2.23	2.23	3.39	2.53	2.40	2.58	2.35	2.53	1.63	2.04	2.09	2.22	2.52	2.61	2.87
Wolof	wol	F	1.47	1.80	2.03	1.99	1.92	1.92	2.10	1.86	1.86	2.25	2.06	1.91	2.10	2.11	1.96	1.61	1.61	1.92	1.85	1.95	2.07	2.03
Xhosa	xho	F	2.09	2.46	2.17	2.90	3.30	3.30	3.60	2.84	2.84	3.93	3.57	3.22	3.62	3.55	2.94	2.18	2.37	2.18	2.93	3.37	3.54	3.57
Yoruba	yor	F	1.57	1.63	1.87	1.95	2.56	2.56	3.16	2.27	2.27	3.32	2.91	2.65	2.94	1.76	2.77	1.54	1.84	1.65	1.70	2.77	3.02	3.02
Zulu	zul	F	2.09	2.48	2.10	3.00	3.37	3.37	3.78	2.91	2.91	4.11	3.74	3.34	3.80	3.72	3.00	2.17	2.38	2.06	3.03	3.49	3.68	3.72
Akan (Asante Twi)	aka	F	1.45	1.65	2.34	2.33	2.23	2.23	2.74	2.09	2.09	2.57	2.73	2.30	2.36	2.38	2.28	1.63	1.59	2.17	2.10	2.35	2.70	2.39
Bambara	bm	F	1.55	2.20	2.59	2.36	2.50	2.50	3.08	2.42	2.42	2.73	3.03	2.51	2.60	2.58	2.46	1.75	1.72	2.17	2.28	2.54	3.00	2.62
Bemba	bem	F	1.77	2.22	2.22	2.34	2.40	2.40	2.63	2.27	2.27	2.90	2.60	2.42	2.64	2.43	2.33	1.83	1.85	2.07	2.28	2.45	2.55	2.61
Central Atlas Tamazight	tzm	W	2.48	12.24	12.32	11.98	5.36	5.36	11.96	11.93	11.93	12.06	11.94	9.02	6.13	2.21	2.12	2.11	2.90	2.19	9.13	10.08	11.97	11.94
Central Kanuri	knc	F	2.18	2.85	2.90	2.83	2.57	2.57	3.07	2.70	2.70	3.17	3.01	2.76	2.83	2.74	2.53	2.23	2.25	2.60	2.60	2.82	2.90	2.95
Dagbani	dag	W	2.06	2.37	2.58	2.46	2.40	2.40	2.67	2.35	2.35	2.77	2.64	2.39	2.52	2.70	2.51	2.02	2.25	2.39	2.23	2.46	2.70	2.45
Deg	mzw	W	1.51	1.69	2.11	1.99	2.09	2.09	2.64	2.02	2.02	2.28	2.64	2.10	2.21	2.07	2.11	1.41	1.55	1.98	1.99	2.17	2.64	2.20
Dinka	din	F	1.61	2.24	2.42	2.30	2.17	2.17	2.60	2.13	2.13	2.48	2.52	2.17	2.37	2.44	2.19	1.66	1.62	2.29	2.11	2.29	2.44	2.25
Dyula	dyu	F	1.80	2.16	2.16	2.16	2.06	2.06	2.30	2.05	2.05	2.41	2.26	2.10	2.22	2.30	2.07	1.79	1.81	2.03	2.03	2.10	2.20	2.26
Edo	bin	W	1.32	1.91	2.01	1.98	2.15	2.15	2.62	2.03	2.03	3.06	2.34	2.22	2.40	2.15	2.29	1.52	1.88	1.68	2.07	2.32	2.40	2.60
Fante	fat	W	1.53	1.76	2.35	2.32	2.26	2.26	2.76	2.14	2.14	2.60	2.75	2.34	2.46	2.47	2.32	1.62	1.69	2.22	2.11	2.38	2.73	2.44
French	fra	F	1.42	2.08	1.60	1.68	1.49	1.49	1.70	1.43	1.43	1.79	1.70	1.44	1.70	2.51	1.56	1.46	1.62	1.43	1.29	1.41	1.43	1.63
Ga	gaa	W	1.60	2.04	2.57	2.46	2.44	2.44	3.38	2.46	2.46	2.80	3.33	2.58	2.68	2.55	2.55	1.62	1.77	2.40	2.28	2.60	3.37	2.67
Ghomala	bbj	W	2.26	2.62	2.85	2.71	2.56	2.56	3.14	2.51	2.51	2.89	3.08	2.59	2.79	2.83	2.52	1.90	2.20	2.79	2.55	2.81	3.05	2.78
Isoko	iso	W	1.57	1.75	1.86	1.87	2.14	2.14	2.51	1.95	1.95	2.83	2.27	2.16	2.30	1.97	2.20	1.48	1.78	1.62	1.90	2.24	2.30	2.47
Kabiye	kbp	F	1.83	3.28	4.36	3.97	3.43	3.43	5.05	4.14	4.14	4.26	5.04	4.23	3.63	3.37	3.59	2.08	1.87	3.29	3.59	4.34	4.96	4.30
Kabuverdianu	kea	F	1.93	2.09	2.10	2.01	1.87	1.87	2.06	1.83	1.83	2.16	2.02	1.85	2.07	2.30	1.82	1.75	1.72	1.95	1.82	1.84	1.91	1.97
Kabyle	kab	F	2.00	2.68	2.94	2.63	2.55	2.55	3.03	2.60	2.60	2.82	2.93	2.65	2.67	2.79	2.55	2.01	2.17	2.64	2.50	2.64	2.90	2.64
Kamba	kam	F	1.77	2.29	2.33	2.37	2.38	2.38	2.79	2.31	2.31	2.80	2.57	2.38	2.59	2.55	2.34	1.85	1.84	2.29	2.28	2.43	2.54	2.58
Kasem	xsm	W	1.45	2.71	3.10	2.84	2.48	2.48	3.27	3.00	3.00	2.99	3.26	3.06	2.51	2.50	2.45	1.67	1.80	2.49	2.74	3.11	3.27	3.11
Kikuyu	kik	F	1.75	2.93	2.93	2.99	2.86	2.86	3.70	2.92	2.92	3.24	3.04	2.90	3.04	3.11	2.94	1.96	1.93	2.91	2.81	2.92	3.02	3.06
Kimbundu	kmb	F	1.76	2.14	2.16	2.21	2.23	2.23	2.53	2.15	2.15	2.74	2.50	2.27	2.54	2.36	2.20	1.77	1.80	2.02	2.17	2.29	2.41	2.47
Kirundi	run	F	1.76	2.14	2.10	2.62	2.65	2.65	2.88	2.37	2.37	3.16	2.85	2.66	2.89	1.98	2.59	1.95	2.01	1.98	2.23	2.70	2.79	2.84
Lingala	lin	F	1.39	1.85	1.88	1.87	1.95	1.95	2.12	1.82	1.82	2.32	2.11	2.01	2.14	2.07	1.95	1.47	1.43	1.81	1.89	2.01	2.08	2.11
Lolak (llg)	llg	W	1.75	1.98	2.11	2.00	1.99	1.99	2.33	2.05	2.05	2.54	2.31	2.14	2.30	2.20	1.94	1.70	1.72	2.03	2.01	2.10	2.22	2.28
Lomwe	ngl	W	2.42	2.66	2.78	2.79	2.84	2.84	3.07	2.74	2.74	3.30	3.05	2.87	3.10	2.94	2.74	2.09	2.48	2.57	2.79	2.86	3.03	3.02
Luganda	lug	F	2.02	2.55	2.58	2.73	2.80	2.80	3.02	2.53	2.53	3.35	2.99	2.81	3.04	2.77	2.70	1.99	2.18	2.38	2.60	2.87	2.95	2.98
Luo	luo	F	1.51	1.91	1.97	1.98	1.91	1.91	2.03	1.85	1.85	2.24	2.02	1.92	2.05	2.16	1.90	1.60	1.55	1.96	1.88	1.94	2.01	2.03
Malagasy	mlg	F	2.01	2.53	2.60	2.61	2.43	2.43	2.65	2.09	2.09	2.86	2.63	2.40	2.67	2.82	2.08	1.64	1.84	1.66	2.45	2.52	2.56	2.55
Mizo	lus	F	1.74	1.88	1.93	1.87	1.72	1.72	1.90	1.72	1.72	2.04	1.85	1.76	1.91	2.19	1.79	1.68	1.42	1.92	1.78	1.77	1.84	1.87
Mooré	mos	F	1.62	2.20	2.48	2.43	2.18	2.18	2.49	2.20	2.20	2.62	2.36	2.24	2.27	2.42	2.18	1.71	1.67	2.26	2.16	2.28	2.33	2.35
Nigerian Fulfulde	fuv	F	1.59	1.76	1.95	1.94	1.99	1.99	2.25	1.92	1.92	2.39	2.22	1.99	2.16	2.03	1.99	1.61	1.67	1.90	2.01	2.10	2.15	2.22
North Ndebele	nde	W	1.97	2.40	1.84	2.71	3.08	3.08	3.41	2.73	2.73	3.73	3.38	3.07	3.42	3.37	2.79	1.97	2.30	2.00	2.83	3.15	3.35	3.36
Northern Sotho	nso	F	1.40	1.67	1.74	2.00	1.90	1.90	2.08	1.76	1.76	2.20	2.05	1.91	2.10	2.20	1.88	1.41	1.45	1.49	1.87	1.94	2.01	2.02
Nuer	nus	F	1.59	2.96	3.40	2.79	2.80	2.80	3.82	3.07	3.07	3.10	3.81	3.13	3.29	2.99	2.83	1.85	2.05	2.82	2.69	3.30	3.70	3.26
Okpe	oke	W	1.52	2.08	2.33	2.27	2.45	2.45	2.98	2.32	2.32	3.44	2.64	2.48	2.73	2.47	2.66	1.74	2.14	2.23	2.29	2.61	2.75	2.89
Sango	sag	F	1.33	1.62	1.61	1.58	1.61	1.61	1.80	1.56	1.56	1.93	1.78	1.65	1.79	1.75	1.61	1.34	1.35	1.58	1.57	1.67	1.73	1.76
South Ndebele	nbl	W	2.10	2.71	2.52	3.00	3.33	3.33	3.66	3.01	3.01	4.00	3.63	3.33	3.68	3.64	3.04	2.27	2.53	2.30	3.11	3.41	3.60	3.60
Swati	ssw	F	2.23	2.91	2.69	3.14	3.33	3.33	3.73	3.08	3.08	4.00	3.69	3.34	3.76	3.73	3.13	2.41	2.48	2.51	3.18	3.42	3.59	3.63
Tamasheq	tmh	F	2.20	2.63	2.82	2.59	2.39	2.39	2.80	2.60	2.60	2.79	2.76	2.54	2.55	2.68	2.43	2.10	2.15	2.56	2.41	2.54	2.71	2.61
Tashelhit	shi	W	1.95	2.46	2.58	2.46	2.21	2.21	2.58	2.34	2.34	2.66	2.52	2.32	2.35	2.52	2.21	1.95	2.02	2.33	2.27	2.35	2.49	2.46
Tiv	tiv	W	1.27	1.72	1.80	1.71	1.73	1.73	1.86	1.70	1.70	2.02	1.85	1.75	1.87	1.90	1.72	1.39	1.57	1.72	1.73	1.81	1.86	1.85
Tongan	ton	W	2.01	2.68	2.45	2.39	2.15	2.15	2.76	1.93	1.93	2.65	2.72	2.14	2.50	2.61	2.31	2.20	2.00	2.40	2.16	2.20	2.39	2.41
Tsonga	tso	F	1.46	1.98	1.95	1.96	2.10	2.10	2.28	1.74	1.74	2.49	2.26	2.11	2.29	2.25	2.03	1.50	1.47	1.84	2.04	2.16	2.23	2.22
Tumbuka	tum	F	1.66	2.14	2.38	2.54	2.54	2.54	2.86	2.31	2.31	3.13	2.84	2.54	2.84	2.62	2.41	1.81	1.79	1.78	2.46	2.65	2.77	2.82
Umbundu	umb	F	2.15	2.70	2.76	2.80	2.82	2.82	3.06	2.66	2.66	3.27	2.99	2.82	3.05	2.99	2.69	2.25	2.20	2.59	2.67	2.84	2.96	3.01
Vagla	vag	W	1.70	1.88	2.40	2.29	2.37	2.37	2.69	2.17	2.17	2.58	2.68	2.24	2.36	2.02	2.03	1.57	1.75	1.95	2.18	2.43	2.67	2.32
Venda	ven	W	1.47	2.14	2.15	2.22	2.22	2.22	2.45	2.01	2.01	2.70	2.41	2.27	2.40	2.36	2.16	1.59	1.84	1.97	2.21	2.29	2.44	2.40