Datasets:

TajikNLPWorld
/

tajik-farsi-transliteration-benchmark

Model stringlengths 16 43	Direction stringclasses 2 values	chrF++_mean float64 10.4 87.4	chrF++_std float64 0.1 6.46	chrF++_ci_low float64 8.64 87.2	chrF++_ci_high float64 13.8 87.4	BLEU_mean float64 0.36 73.6	BLEU_std float64 0.09 7.68	BLEU_ci_low float64 0.1 73.3	BLEU_ci_high float64 0.45 73.7	TER_mean float64 17.2 215	TER_std float64 0.08 37.3	TER_ci_low float64 17.1 193	TER_ci_high float64 17.3 268	CER_mean float64 0.05 2.45	CER_std float64 0 0.4	CER_ci_low float64 0.05 2.35	CER_ci_high float64 0.06 3.01	WER_mean float64 0.21 3.34	WER_std float64 0 0.59	WER_ci_low float64 0.21 3.17	WER_ci_high float64 0.22 4.14	Acc%_mean float64 0 51.1	Acc%_std float64 0 7.51	Acc%_ci_low float64 0 50.3	Acc%_ci_high float64 0 51.8	train_time_s float64 0 2.98k	infer_time_ms float64 0 332 ⌀	gpu_mem_gb float64 0 18 ⌀
Baseline_farsi2tajik	farsi2tajik	10.401701	2.43727	8.637874	13.848202	0.359565	0.086466	0.241463	0.446068	107.897913	0.422237	107.307992	108.273005	0.537021	0.076089	0.429415	0.590841	1.131019	0.005823	1.12286	1.13606	0.041667	0.011785	0.025	0.05	0	0	0
Baseline_tajik2farsi	tajik2farsi	13.99143	6.036195	9.707939	22.52788	0.415308	0.191309	0.230304	0.678774	96.53473	2.042915	93.645644	97.991098	0.616329	0.141253	0.416569	0.716931	0.970786	0.023028	0.93822	0.987299	0.458333	0.418496	0.15	1.05	0	0	0
CharTransformer_farsi2tajik	farsi2tajik	18.234791	1.762826	16.068921	20.386878	0.446911	0.255453	0.145755	0.770299	207.969013	15.197613	192.57741	228.656161	2.449557	0.125232	2.353929	2.626468	3.336087	0.234781	3.168157	3.668111	0	0	0	0	1,699.744632	null	null
CharTransformer_tajik2farsi	tajik2farsi	17.910574	1.535164	16.236859	19.944972	0.389319	0.275736	0.096324	0.758664	215.100472	37.252554	184.694655	267.562525	2.449622	0.40089	2.106643	3.012063	3.307715	0.592196	2.784846	4.135719	0	0	0	0	1,633.967302	null	null
G2PTransformer_farsi2tajik	farsi2tajik	60.266954	0.742248	59.219067	60.84425	21.00309	1.357501	19.451918	22.758275	60.491585	4.784976	56.69755	67.241228	0.473803	0.045595	0.426926	0.535586	0.741418	0.05152	0.684763	0.80942	0	0	0	0	1,206.716265	null	null
G2PTransformer_tajik2farsi	tajik2farsi	72.261975	0.409927	71.682254	72.553704	36.461921	0.407476	35.890944	36.814815	39.608888	1.051685	38.717745	41.085701	0.41546	0.020188	0.399228	0.443915	0.50798	0.02149	0.489186	0.538061	0	0	0	0	1,149.048677	null	null
LSTM_farsi2tajik	farsi2tajik	31.445293	6.460367	24.372483	39.990295	4.943294	3.391382	1.566276	9.581213	113.307921	22.368687	84.126286	138.475193	0.358739	0.075356	0.262185	0.446077	0.960512	0.128566	0.801481	1.116354	1.791667	1.11455	0.225	2.725	2,298.995266	null	null
LSTM_tajik2farsi	tajik2farsi	65.104811	5.512486	57.309064	69.034436	38.521127	7.678316	27.663302	44.074699	52.047873	13.791667	41.701752	71.539922	0.136873	0.027682	0.11627	0.176003	0.442961	0.061748	0.395549	0.530175	23.825	3.713882	18.6	26.9	2,983.459352	null	null
byt5-small_farsi2tajik	farsi2tajik	80.070879	0.227039	79.79212	80.348244	56.607236	0.320455	56.315953	57.053549	28.173091	0.222522	27.899814	28.444873	0.090392	0.001032	0.089307	0.091779	0.359274	0.003314	0.355236	0.363353	23.008333	0.305732	22.7	23.425	2,344.033252	246.679552	9.353318
byt5-small_tajik2farsi	tajik2farsi	87.35401	0.095886	87.218628	87.428406	73.57503	0.1797	73.326311	73.744584	17.247947	0.078825	17.145528	17.337271	0.054277	0.000597	0.053786	0.055117	0.214761	0.001317	0.213545	0.216591	51.066667	0.633224	50.275	51.825	2,430.106475	223.940571	9.353871
mbart-large-50-many-to-many-mmt_farsi2tajik	farsi2tajik	70.110926	0.442553	69.531632	70.60574	45.385396	0.395863	44.982494	45.923468	43.80097	0.909755	42.567392	44.734295	0.23829	0.006712	0.232692	0.247727	0.599907	0.006837	0.591711	0.608447	0.85	0.201039	0.575	1.05	2,606.910462	163.267521	17.998621
mbart-large-50-many-to-many-mmt_tajik2farsi	tajik2farsi	62.165485	5.33346	54.675433	66.680601	44.204779	6.277315	35.626936	50.474225	42.214551	5.078373	38.337544	49.388646	0.382058	0.067852	0.309744	0.47284	0.533778	0.058781	0.474142	0.613751	5.425	7.513072	0.075	16.05	2,471.656167	80.95754	16.19186
mt5-small_farsi2tajik	farsi2tajik	14.264789	0.201852	13.980937	14.432926	1.747083	0.09462	1.617758	1.841507	130.911986	3.759144	126.612694	135.769734	0.754385	0.015054	0.739223	0.774909	1.13835	0.027445	1.107961	1.174455	0.033333	0.03118	0	0.075	1,777.365384	331.955854	10.656398
mt5-small_tajik2farsi	tajik2farsi	18.487018	0.919127	17.749478	19.782729	3.449899	0.667035	2.888786	4.387166	115.763215	12.540404	101.325923	131.90184	0.749872	0.048291	0.687909	0.805722	1.080736	0.098798	0.974317	1.212353	0.075	0.054006	0	0.125	1,774.334839	154.692658	10.459706

YAML Metadata Warning:The task_categories "text2text-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other

🇹🇯🇮🇷 Tajik-Farsi Transliteration Benchmark

Официальный бенчмарк машинной транслитерации между таджикским (кириллица) и фарси (персо-арабская графика).
Результаты получены на 40k параллельных предложениях с оценкой по 3 случайным сидам, bootstrap 95% CI и парными статистическими тестами.

📊 Ключевые результаты (Top-5)

Модель	Направление	chrF++	BLEU	CER
byt5-small	Tj→Fa	87.35 ± 0.10	73.58	0.054
byt5-small	Fa→Tj	80.07 ± 0.23	56.61	0.090
G2PTransformer	Tj→Fa	72.26 ± 0.41	36.46	0.415
mbart-large-50-many-to-many-mmt	Fa→Tj	70.11 ± 0.44	45.39	0.238
LSTM	Tj→Fa	65.10 ± 5.51	38.52	0.137

Примечание: ByT5-small демонстрирует наивысшую стабильность (σ < 0.25). G2P-Transformer превосходит mBART/mT5 в направлении Tj→Fa при ~10× меньшем числе параметров.

📁 Структура репозитория

📦 tajik-farsi-transliteration-benchmark/
├── 📄 README.md                 ← Этот файл
├── 📊 results/
│   ├── aggregated_metrics.csv   ← Сводная таблица метрик
│   ├── statistical_report.json  ← p-значения, ранги, CI
│   └── inference_samples.json   ← Примеры предсказаний
└── 📈 plots/
    ├── pareto_frontier.png      ← Качество vs. время обучения
    └── interactive_report.html  ← HTML-отчёт с таблицами

🛠 Как использовать результаты

import pandas as pd
from datasets import load_dataset

# Загрузить метрики
ds = load_dataset("TajikNLPWorld/tajik-farsi-transliteration-benchmark", split="train")
df = ds.to_pandas()
print(df.sort_values("chrF++_mean", ascending=False).head())

📖 Citation

@misc{tajikfarsi_benchmark_2026,
  author = {[Ваше Имя] и соавторы},
  title = {Tajik-Farsi Transliteration Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/TajikNLPWorld/tajik-farsi-transliteration-benchmark}
}

⚖️ License

MIT License. Код и данные открыты для исследовательского и коммерческого использования.

Downloads last month: 28

Number of rows:

Total file size:

123 kB