Datasets:

dimhold
/

tokenizer-detect

lang stringlengths 8 8	tokenizer stringclasses 4 values	sentences int64 1.01k 1.01k	words int64 1.08k 36.1k	chars int64 39.4k 179k	tokens int64 21.5k 506k	tokens_per_word float64 1.17 197	tokens_per_char float64 0.15 3.16	single_char_token_share float64 0.04 1
ace_Arab	cl100k_base	1,012	21,863	90,770	102,566	4.691305	1.129955	0.872326
ace_Arab	o200k_base	1,012	21,863	90,770	70,310	3.215936	0.774595	0.605718
ace_Arab	gpt2	1,012	21,863	90,770	129,388	5.918127	1.425449	0.898476
ace_Arab	xlm-roberta-base	1,012	21,863	90,770	59,462	2.719755	0.655084	0.312166
ace_Latn	cl100k_base	1,012	21,328	121,617	54,216	2.542011	0.445793	0.221909
ace_Latn	o200k_base	1,012	21,328	121,617	48,788	2.287509	0.401161	0.171477
ace_Latn	gpt2	1,012	21,328	121,617	59,798	2.803732	0.491691	0.302335
ace_Latn	xlm-roberta-base	1,012	21,328	121,617	47,988	2.25	0.394583	0.212136
acm_Arab	cl100k_base	1,012	18,733	95,661	80,923	4.31981	0.845935	0.806013
acm_Arab	o200k_base	1,012	18,733	95,661	37,105	1.980729	0.38788	0.206172
acm_Arab	gpt2	1,012	18,733	95,661	115,076	6.142956	1.202956	0.895851
acm_Arab	xlm-roberta-base	1,012	18,733	95,661	35,415	1.890514	0.370214	0.22129
acq_Arab	cl100k_base	1,012	19,234	96,891	81,638	4.244463	0.842576	0.805005
acq_Arab	o200k_base	1,012	19,234	96,891	37,357	1.942238	0.385557	0.212544
acq_Arab	gpt2	1,012	19,234	96,891	117,377	6.102579	1.211433	0.901292
acq_Arab	xlm-roberta-base	1,012	19,234	96,891	35,927	1.86789	0.370798	0.226932
aeb_Arab	cl100k_base	1,012	18,828	94,176	79,190	4.20597	0.840872	0.803561
aeb_Arab	o200k_base	1,012	18,828	94,176	38,188	2.028256	0.405496	0.224233
aeb_Arab	gpt2	1,012	18,828	94,176	113,473	6.026822	1.204904	0.899104
aeb_Arab	xlm-roberta-base	1,012	18,828	94,176	36,709	1.949703	0.389791	0.236454
afr_Latn	cl100k_base	1,012	23,082	117,620	45,906	1.988822	0.390291	0.183636
afr_Latn	o200k_base	1,012	23,082	117,620	38,367	1.662204	0.326195	0.141163
afr_Latn	gpt2	1,012	23,082	117,620	53,316	2.309852	0.45329	0.255833
afr_Latn	xlm-roberta-base	1,012	23,082	117,620	36,694	1.589724	0.311971	0.170328
ajp_Arab	cl100k_base	1,012	18,890	91,168	76,997	4.076072	0.844562	0.80607
ajp_Arab	o200k_base	1,012	18,890	91,168	36,027	1.9072	0.395172	0.217948
ajp_Arab	gpt2	1,012	18,890	91,168	108,854	5.76252	1.193994	0.895833
ajp_Arab	xlm-roberta-base	1,012	18,890	91,168	34,361	1.819005	0.376898	0.212246
aka_Latn	cl100k_base	1,012	26,781	106,881	73,267	2.735783	0.685501	0.556076
aka_Latn	o200k_base	1,012	26,781	106,881	56,050	2.092902	0.524415	0.397484
aka_Latn	gpt2	1,012	26,781	106,881	77,062	2.877488	0.721007	0.583764
aka_Latn	xlm-roberta-base	1,012	26,781	106,881	61,023	2.278593	0.570943	0.441358
als_Latn	cl100k_base	1,012	24,611	123,465	61,065	2.481208	0.494594	0.366446
als_Latn	o200k_base	1,012	24,611	123,465	49,215	1.999716	0.398615	0.208737
als_Latn	gpt2	1,012	24,611	123,465	72,324	2.938686	0.585785	0.535479
als_Latn	xlm-roberta-base	1,012	24,611	123,465	40,383	1.640852	0.327081	0.204492
amh_Ethi	cl100k_base	1,012	17,330	70,814	207,397	11.967513	2.928757	0.982555
amh_Ethi	o200k_base	1,012	17,330	70,814	156,283	9.018061	2.206951	0.874292
amh_Ethi	gpt2	1,012	17,330	70,814	209,308	12.077784	2.955743	0.993221
amh_Ethi	xlm-roberta-base	1,012	17,330	70,814	40,460	2.33468	0.571356	0.455635
apc_Arab	cl100k_base	1,012	17,898	91,094	76,577	4.278523	0.840637	0.81087
apc_Arab	o200k_base	1,012	17,898	91,094	36,164	2.020561	0.396997	0.213887
apc_Arab	gpt2	1,012	17,898	91,094	108,981	6.089004	1.196358	0.900533
apc_Arab	xlm-roberta-base	1,012	17,898	91,094	35,084	1.960219	0.385141	0.221982
arb_Arab	cl100k_base	1,012	19,637	97,682	82,296	4.190864	0.842489	0.801533
arb_Arab	o200k_base	1,012	19,637	97,682	37,123	1.890462	0.380039	0.206395
arb_Arab	gpt2	1,012	19,637	97,682	118,886	6.054183	1.217072	0.900342
arb_Arab	xlm-roberta-base	1,012	19,637	97,682	35,967	1.831593	0.368205	0.230461
arb_Latn	cl100k_base	1,012	22,179	132,570	66,596	3.00266	0.502346	0.280212
arb_Latn	o200k_base	1,012	22,179	132,570	60,552	2.73015	0.456755	0.234394
arb_Latn	gpt2	1,012	22,179	132,570	68,990	3.1106	0.520404	0.337846
arb_Latn	xlm-roberta-base	1,012	22,179	132,570	59,263	2.672032	0.447032	0.284798
ars_Arab	cl100k_base	1,012	19,656	97,784	82,423	4.193274	0.842909	0.801912
ars_Arab	o200k_base	1,012	19,656	97,784	37,240	1.894587	0.380839	0.207358
ars_Arab	gpt2	1,012	19,656	97,784	119,047	6.056522	1.217449	0.900375
ars_Arab	xlm-roberta-base	1,012	19,656	97,784	36,166	1.839947	0.369856	0.231986
ary_Arab	cl100k_base	1,012	19,117	95,457	80,399	4.205628	0.842254	0.806291
ary_Arab	o200k_base	1,012	19,117	95,457	39,862	2.08516	0.417591	0.231097
ary_Arab	gpt2	1,012	19,117	95,457	114,026	5.964639	1.194527	0.890078
ary_Arab	xlm-roberta-base	1,012	19,117	95,457	38,046	1.990166	0.398567	0.214267
arz_Arab	cl100k_base	1,012	19,505	95,176	80,253	4.114483	0.843206	0.806562
arz_Arab	o200k_base	1,012	19,505	95,176	37,779	1.936888	0.396938	0.205273
arz_Arab	gpt2	1,012	19,505	95,176	114,490	5.869777	1.202929	0.901258
arz_Arab	xlm-roberta-base	1,012	19,505	95,176	35,818	1.83635	0.376334	0.193311
asm_Beng	cl100k_base	1,012	19,289	107,564	168,150	8.717404	1.563255	0.897698
asm_Beng	o200k_base	1,012	19,289	107,564	53,545	2.775934	0.497797	0.368979
asm_Beng	gpt2	1,012	19,289	107,564	262,509	13.609259	2.440491	0.925972
asm_Beng	xlm-roberta-base	1,012	19,289	107,564	57,803	2.996682	0.537382	0.40036
ast_Latn	cl100k_base	1,012	21,954	115,421	43,202	1.967842	0.374299	0.243878
ast_Latn	o200k_base	1,012	21,954	115,421	38,261	1.74278	0.331491	0.204961
ast_Latn	gpt2	1,012	21,954	115,421	52,330	2.38362	0.453384	0.311313
ast_Latn	xlm-roberta-base	1,012	21,954	115,421	39,125	1.782135	0.338976	0.230058
awa_Deva	cl100k_base	1,012	25,403	103,650	129,962	5.11601	1.253854	0.865807
awa_Deva	o200k_base	1,012	25,403	103,650	45,866	1.805535	0.442508	0.30092
awa_Deva	gpt2	1,012	25,403	103,650	194,354	7.650829	1.875099	0.991937
awa_Deva	xlm-roberta-base	1,012	25,403	103,650	41,850	1.647443	0.403763	0.261673
ayr_Latn	cl100k_base	1,012	16,000	123,629	59,171	3.698187	0.478617	0.301499
ayr_Latn	o200k_base	1,012	16,000	123,629	52,125	3.257812	0.421624	0.242283
ayr_Latn	gpt2	1,012	16,000	123,629	64,075	4.004688	0.518285	0.360031
ayr_Latn	xlm-roberta-base	1,012	16,000	123,629	52,077	3.254812	0.421236	0.264704
azb_Arab	cl100k_base	1,012	17,603	100,464	90,306	5.130148	0.898889	0.817199
azb_Arab	o200k_base	1,012	17,603	100,464	49,962	2.838266	0.497312	0.323446
azb_Arab	gpt2	1,012	17,603	100,464	139,579	7.929273	1.389343	0.915102
azb_Arab	xlm-roberta-base	1,012	17,603	100,464	43,458	2.468784	0.432573	0.222284
azj_Latn	cl100k_base	1,012	19,105	125,646	71,733	3.754672	0.570914	0.502028
azj_Latn	o200k_base	1,012	19,105	125,646	46,080	2.411934	0.366745	0.198893
azj_Latn	gpt2	1,012	19,105	125,646	94,631	4.953206	0.753156	0.650294
azj_Latn	xlm-roberta-base	1,012	19,105	125,646	35,170	1.840879	0.279913	0.164316
bak_Cyrl	cl100k_base	1,012	17,949	116,279	115,692	6.445596	0.994952	0.780797
bak_Cyrl	o200k_base	1,012	17,949	116,279	46,911	2.613572	0.403435	0.247383
bak_Cyrl	gpt2	1,012	17,949	116,279	163,052	9.084183	1.402248	0.939762
bak_Cyrl	xlm-roberta-base	1,012	17,949	116,279	62,884	3.503482	0.540803	0.428503
bam_Latn	cl100k_base	1,012	22,566	104,665	69,577	3.083267	0.664759	0.57325
bam_Latn	o200k_base	1,012	22,566	104,665	54,799	2.428388	0.523566	0.443566
bam_Latn	gpt2	1,012	22,566	104,665	72,894	3.230258	0.696451	0.585906
bam_Latn	xlm-roberta-base	1,012	22,566	104,665	55,516	2.460161	0.530416	0.487409
ban_Latn	cl100k_base	1,012	20,343	126,201	48,715	2.394681	0.386011	0.180519
ban_Latn	o200k_base	1,012	20,343	126,201	42,283	2.078504	0.335045	0.142563
ban_Latn	gpt2	1,012	20,343	126,201	54,141	2.661407	0.429006	0.23062
ban_Latn	xlm-roberta-base	1,012	20,343	126,201	39,983	1.965443	0.31682	0.137383

End of preview. Expand in Data Studio

tokenizer-detect

Author: Dmitriy Semenkevich (dimhold)

This dataset holds tokenizer fertility (tokens per whitespace word) over FLORES-200 devtest for 204 languages and 4 tokenizers (cl100k_base, o200k_base, gpt2, xlm-roberta-base), the X axis of a test of whether fertility explains machine text detector quality.

Headline, from the repository: the answer to that question is "No": "the sign of the correlation is set by the language the detector was trained on", with partial rho -0.884 against fertility for MULTITuDE detectors fine tuned on English and +0.747 for those fine tuned on Russian.

This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.

Code and full write up: https://github.com/dimhold/tokenizer-detect
DOI: https://doi.org/10.5281/zenodo.22417816

Files

path	what it is
`fertility.tsv`	816 rows, 204 languages x 4 tokenizers: sentences, words, characters, tokens, tokens per word, tokens per character, single character token share

How it was produced

Counted by fertility.py (in the repository) over the 1012 FLORES-200 devtest sentences; this file holds counts only, no sentence text. Not mirrored here: the detector quality tables copied from MULTITuDE and Pangram 4, the CommonCrawl language shares and the join outputs, since they reproduce third party published tables. Get them from the GitHub repository. The scripts that turn these files into every number are in the GitHub repository.

Licence

MIT, the licence of the repository, covering code and data.

Citation

@misc{semenkevich2026tokenizerdetect,
  author = {Semenkevich, Dmitriy},
  title  = {tokenizer-detect: tokenizer fertility over 204 languages},
  year   = {2026},
  doi    = {10.5281/zenodo.22417816},
  url    = {https://doi.org/10.5281/zenodo.22417816}
}

Downloads last month: -