Datasets:

NagaYu
/

parity-fertility-atlas

lang stringlengths 2 2	tokenizer_id stringclasses 2 values	n_sentences int64 2k 2k	total_tokens int64 26.3k 353k	total_chars int64 45k 190k	total_words int64 9.6k 39.4k ⌀	tokens_per_char float64 0.3 2.86	tokens_per_word float64 1.29 17.3 ⌀	parity_ratio float64 0.92 13.5 ⌀	parity_ratio_median float64 0.91 14 ⌀	effective_context_fraction float64 0.07 1.09 ⌀	corpus stringclasses 1 value	split stringclasses 1 value
ja	HuggingFaceTB/SmolLM2-135M	2,000	79,648	44,969	null	1.771176	null	2.981285	3	0.335426	opus100	devtest
hi	HuggingFaceTB/SmolLM2-135M	2,000	165,529	122,316	31,564	1.35329	5.244234	4.911403	4.875	0.203608	opus100	devtest
ar	HuggingFaceTB/SmolLM2-135M	2,000	92,551	81,212	17,448	1.139622	5.30439	3.322599	3.077677	0.300969	opus100	devtest
th	HuggingFaceTB/SmolLM2-135M	2,000	125,792	66,830	null	1.882268	null	6.230103	6	0.160511	opus100	devtest
ko	HuggingFaceTB/SmolLM2-135M	2,000	119,010	46,635	15,572	2.551946	7.642564	4.101106	4.1	0.243837	opus100	devtest
bn	HuggingFaceTB/SmolLM2-135M	2,000	303,794	138,104	39,416	2.199748	7.707378	8.502015	8.166667	0.117619	opus100	devtest
te	HuggingFaceTB/SmolLM2-135M	2,000	166,360	58,398	9,602	2.848728	17.325557	10.041043	10.142857	0.099591	opus100	devtest
ta	HuggingFaceTB/SmolLM2-135M	2,000	353,480	189,997	22,926	1.86045	15.418302	9.571363	9.226496	0.104478	opus100	devtest
am	HuggingFaceTB/SmolLM2-135M	2,000	311,226	108,934	33,372	2.857014	9.325962	6.570524	6.427521	0.152195	opus100	devtest
my	HuggingFaceTB/SmolLM2-135M	2,000	237,212	83,568	null	2.838551	null	13.507118	14	0.074035	opus100	devtest
zh	HuggingFaceTB/SmolLM2-135M	2,000	117,622	67,913	null	1.731951	null	2.370694	2.353394	0.421817	opus100	devtest
ru	HuggingFaceTB/SmolLM2-135M	2,000	68,476	76,106	13,997	0.899745	4.892191	3.123193	2.866667	0.320185	opus100	devtest
es	HuggingFaceTB/SmolLM2-135M	2,000	38,908	88,926	17,945	0.437532	2.168181	1.601218	1.527047	0.624525	opus100	devtest
en	HuggingFaceTB/SmolLM2-135M	2,000	26,716	89,142	20,359	0.299702	1.312245	null	null	null	opus100	devtest
ja	Qwen/Qwen2.5-0.5B-Instruct	2,000	32,089	44,969	null	0.71358	null	1.219326	1.175192	0.820125	opus100	devtest
hi	Qwen/Qwen2.5-0.5B-Instruct	2,000	141,812	122,316	31,564	1.15939	4.49284	4.265664	4.25	0.23443	opus100	devtest
ar	Qwen/Qwen2.5-0.5B-Instruct	2,000	38,519	81,212	17,448	0.474302	2.207646	1.402527	1.333333	0.712999	opus100	devtest
th	Qwen/Qwen2.5-0.5B-Instruct	2,000	39,610	66,830	null	0.592698	null	1.992555	1.888889	0.501868	opus100	devtest
ko	Qwen/Qwen2.5-0.5B-Instruct	2,000	40,547	46,635	15,572	0.869454	2.60384	1.414019	1.428571	0.707204	opus100	devtest
bn	Qwen/Qwen2.5-0.5B-Instruct	2,000	180,652	138,104	39,416	1.308087	4.583215	5.099562	4.850384	0.196095	opus100	devtest
te	Qwen/Qwen2.5-0.5B-Instruct	2,000	92,930	58,398	9,602	1.591322	9.678192	5.832183	5.75	0.171462	opus100	devtest
ta	Qwen/Qwen2.5-0.5B-Instruct	2,000	229,254	189,997	22,926	1.206619	9.999738	6.284892	6.1	0.159112	opus100	devtest
am	Qwen/Qwen2.5-0.5B-Instruct	2,000	177,089	108,934	33,372	1.625654	5.306514	3.765768	3.642857	0.26555	opus100	devtest
my	Qwen/Qwen2.5-0.5B-Instruct	2,000	128,673	83,568	null	1.53974	null	7.70082	8.211111	0.129856	opus100	devtest
zh	Qwen/Qwen2.5-0.5B-Instruct	2,000	45,060	67,913	null	0.663496	null	0.919911	0.905084	1.087062	opus100	devtest
ru	Qwen/Qwen2.5-0.5B-Instruct	2,000	33,083	76,106	13,997	0.434696	2.363578	1.534818	1.444444	0.651543	opus100	devtest
es	Qwen/Qwen2.5-0.5B-Instruct	2,000	30,519	88,926	17,945	0.343195	1.700697	1.280159	1.237798	0.781153	opus100	devtest
en	Qwen/Qwen2.5-0.5B-Instruct	2,000	26,317	89,142	20,359	0.295226	1.292647	null	null	null	opus100	devtest

Parity fertility atlas

How many tokens each tokenizer charges for the same meaning, measured on a parallel corpus (opus100).

column	meaning
`tokenizer_id`	the tokenizer measured
`lang`	ISO code
`tokens_per_char`	tokens per NFC character, excluding whitespace
`tokens_per_word`	tokens per whitespace word; `null` for scripts without word spaces
`parity_ratio`	tokens(target) / tokens(aligned English) — the headline
`parity_ratio_median`	median of the per-sentence ratios
`effective_context_fraction`	`1 / parity_ratio`: the share of the advertised context window a speaker of this language actually receives

How to read this

A parity_ratio of 2.4 means a speaker of that language spends 2.4x the tokens to say the same thing, gets 42% of the context window, and pays 2.4x per message.

This is a property of the tokenizer, not of the language. A tokenizer is fitted to a corpus; the corpora used to fit today's tokenizers under-represent most of the world's writing systems, and the ratios below are the arithmetic consequence. Nothing here says any language is inefficient, verbose, or harder to model. The measurement is of an engineering artefact, and the artefact is repairable — that is what the Parity repository is for.

Reproducing

python scripts/build_atlas.py --tokenizers HuggingFaceTB/SmolLM2-135M,Qwen/Qwen2.5-0.5B-Instruct --langs am,ar,bn,en,es,hi,ja,ko,my,ru,ta,te,th,zh

Rows are produced by parity.fertility.measure_fertility on 2000 aligned sentences from opus100.

Web-mined corpora are cleaned first (parity.corpora.clean_bitext) — badly aligned pairs otherwise dominate a corpus-level ratio. Both the corpus ratio and the per-sentence median are given so surviving noise is visible rather than implied.

The repair

A high ratio here is fixable without retraining the model: see parity build and the demo Space. Measurement is the smallest part of that project; the reduction and its certified bound are the point.

Downloads last month: 24

Number of rows:

Total file size:

12.6 kB

Parity fertility atlas

How to read this

Reproducing

The repair

Space using NagaYu/parity-fertility-atlas 1