PoetryMTEB/MultilingualPoetryBitextMining

Multilingual Poetry Bitext Mining

Multilingual poetry translation bitexts for PoetryMTEB / MTEB-style bitext mining evaluation, covering 23 directed language pairs (10,943 pairs).

Pairs are constructed by structurally extracting bilingual poetry alignments from lyrikline using an LLM (DeepSeek V3.1). Each Hub config is one directed language pair (source → target).

Access & copyright. This dataset includes poems and translations that cannot be openly licensed. It is not for public redistribution. Request access on the Hub if you need to use it for research evaluation. See License & access.

Dataset Card

Item	Description
Dataset version (PoetryMTEB)	1.1.2
Hub repo	PoetryMTEB/MultilingualPoetryBitextMining
Task	Bitext mining / parallel text retrieval
Unit	Aligned source_text ↔ target_text pair
Configs	23 language pairs (eng-deu default)
Total pairs	10,943
Min pairs / config	125
Primary source	https://www.lyrikline.org/
Extraction model	DeepSeek V3.1
Access	Gated (manual approval) — request access on the Hub
License	other (poetrymteb-mixed-copyright-research-use)
Intended use	Restricted non-commercial research evaluation (after approval)

Collections (configs)

Config	Direction	n
bel-deu	Belarusian → German	190
cat-deu	Catalan → German	451
cat-eng	Catalan → English	623
cat-spa	Catalan → Spanish	206
deu-eng	German → English	675
deu-fra	German → French	532
deu-ita	German → Italian	471
deu-nld	German → Dutch	507
deu-rus	German → Russian	125
deu-spa	German → Spanish	542
eng-deu	English → German	1,077
fas-deu	Persian → German	198
fra-deu	French → German	632
hun-deu	Hungarian → German	558
ita-deu	Italian → German	445
lit-deu	Lithuanian → German	579
nld-deu	Dutch → German	639
nld-eng	Dutch → English	175
por-deu	Portuguese → German	409
rus-deu	Russian → German	338
slv-deu	Slovenian → German	716
spa-deu	Spanish → German	700
tur-deu	Turkish → German	155

Total: 10,943 pairs across 23 configs.

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/MultilingualPoetryBitextMining", "eng-deu", split="test")
print(len(ds), ds[0]["id"])
print(ds[0]["source_text"][:120])
print(ds[0]["target_text"][:120])

Features

Field	Type	Description
id	string	Pair identifier (often {poem_hash}_{line_or_span_index})
source_text	string	Text in the source language of the config name
target_text	string	Aligned text in the target language

Construction overview

lyrikline.org bilingual poetry pages
        │
        ▼
  DeepSeek V3.1 structural extraction of bilingual poetry pairs
        │
        ▼
  23 directed language-pair configs (test split only)

Bitexts are structurally extracted from lyrikline.org with DeepSeek V3.1 (10,943 pairs across 23 configs).

License & access

This dataset is not openly redistributable.

It contains poems and translations from lyrikline that remain under copyright (contemporary poets and translators). lyrikline “free of charge” means free website access, not an open license for third-party redistribution. This Hub packaging is therefore:

1. Not released under Creative Commons (or any open-content license) for the poem/translation texts.
2. Gated — you must request access on Hugging Face and wait for approval before downloading.
3. Non-redistribution — approved access is for non-commercial research / PoetryMTEB evaluation only; do not mirror, republish, or otherwise publicly redistribute the texts. Approval is not a copyright license.
4. Rights remain with the poets, translators, and upstream rights holders. Takedown requests: Hub discussion / PoetryMTEB contact channels.

License tag: other · Name: poetrymteb-mixed-copyright-research-use · See LICENSE.

Intended use

• PoetryMTEB multilingual bitext-mining evaluation after Hub access approval.
• Research use only; no public redistribution of the underlying texts.

Limitations

• Alignments are LLM-extracted; residual noise may remain.
• Pair counts are unbalanced across configs (≈125–1077).
• Many configs are deu-centric, reflecting lyrikline coverage in this extraction.
• Only a test split is released.
• Copyright-restricted content: gated access required.

Citation

@misc{multilingualpoetrybitextmining,
  title     = {Multilingual Poetry Bitext Mining},
  author    = {PoetryMTEB Contributors},
  year      = {2026},
  version   = {1.1.2},
  url       = {https://huggingface.co/datasets/PoetryMTEB/MultilingualPoetryBitextMining},
  note      = {PoetryMTEB packaging; lyrikline bitexts; DeepSeek V3.1 extraction; gated mixed copyright}
}

Please attribute lyrikline (https://www.lyrikline.org/) and the original poets / translators when publishing results that depend on these texts.

Hub packaging: PoetryMTEB/MultilingualPoetryBitextMining (version 1.1.2); total pairs=10,943; parquet bytes=2,882,028.