CrocoSign/pjm-data

PJM-to-text dataset

Sentence-level Polish Sign Language (PJM) video paired with Polish transcripts and English translations, ready for training sign language translation models. 29,928 video-sentence pairs, about 25 hours of video, from 151 Deaf signers, with a multi-signer and a signer-independent split.

We built it from the Open Repository of the PJM Corpus of the Sign Language Linguistics Lab, University of Warsaw: we cut the recordings into one clip per subtitle sentence, corrected the Polish transcripts, translated them to English, identified the signer of every clip and filtered the clips for quality. The method and the baseline models are described in our paper, A Polish Sign Language to Text Dataset and Benchmark (in preparation). Overview: crocosign.github.io.

Related repositories:

• CrocoSign/pjm-segments: all 39,817 sentence clips before filtering, as WebDataset shards.
• CrocoSign/pjm-models: the trained models, their checkpoints and test outputs.
• github.com/crocosign/pjm-data: the code that built this dataset.

Contents

Folder or file	What it holds
splits/	The pair lists. See Splits below. pjm_data.csv is the metadata of every source recording (URLs, video type, task, age group, sex, signer id).
texts/texts_fixed.csv	Polish and English text of the 29,928 pairs, as first published and corrected (see texts/README.md).
vectors/	Pre-extracted features in HDF5, one dataset per clip key: vit_feat_pjm.h5 (CLIP ViT-L/14), mae_feat_pjm.h5 (VideoMAE-large), hand_vit_feat_pjm.h5 (ViT on hand crops), mediapipe_feat_pjm.h5 (MediaPipe keypoints), texts.h5 (Polish, without diacritics), texts_eng.h5 (English), texts_multilang_pjm.h5 (Polish, English, French, Spanish).
poses/	RTMPose COCO-WholeBody keypoints of every clip, in tar shards (see poses/README.md).
crop_params/	Per-clip square crop around the signer, as LMDB (see crop_params/README.md).
videos/	The two single-signer halves of each of the 198 dialogue recordings (396 files). Monologues are used as published by the corpus, at the URLs in splits/pjm_data.csv.
translations/	The Polish WebVTT subtitles of the 730 single-signer videos.
dialogs_data.csv	Metadata of the split dialogue videos.

The sentence clips themselves are in pjm-segments; a clip key such as 004_000_M is {recording}_{sentence}_{side}, where the side is L or R for a dialogue half and M for a monologue. The key joins the split files, the features, the poses and the clips.

Splits

Split	Files	Train / validation / test
Multi-signer (MS)	split_{train,val,test}_ms.csv	23,941 / 3,000 / 2,987
Signer-independent (SI)	split_{train,val,test}_filtered.csv	10,062 / 9,956 / 9,910

Both splits cover the same 29,928 pairs. In MS, 80%, 10% and 10% of each signer's clips go to train, validation and test, so test signers also appear in training. In SI, every signer is in exactly one set (51, 50 and 50 signers). About 30% of the distinct words in the SI validation and test sets do not occur in SI training, against about 10% for MS.

split_{train,val,test}.csv is the signer-independent split before the final quality filter (33,897 pairs). We kept it because two baseline rows in the paper were trained on it; use the _filtered files for new work.

Selection

From 39,817 sentence clips, five filters removed translations flagged as invalid, empty or numeric sentences, one-word sentences, sentences dominated by rare words (seen at most three times in the corpus) and sentences longer than 60 sub-word tokens, leaving 33,897. A final quality filter kept sentences with at least three words, at least one word outside the most common ones and a valid translation, leaving 29,932. Four of them have broken video (one to five frames) and were removed, which leaves 29,928.

Known limitations

• The English references are DeepL translations of the Polish transcripts. A later check found meaning errors in 1,440 of them (4.8%), mostly systematic mistranslations of a few words, such as migać (to sign) rendered as "blink". texts/texts_fixed.csv holds the corrected texts next to the originals; the results in our paper use the originals.
• Each sentence has one reference, and many valid renderings exist.
• About 25 hours from 151 signers is small for sign language translation, and every signer contributes little video.
• The signer identifiers come from face clustering with manual checks; they encode sex, age group and a running number, not names.

Licence and attribution

The corpus is released under CC BY-SA 4.0, and so is this dataset. The recordings show the faces of the signers who recorded themselves for the corpus; please treat them with respect. When you use the dataset, cite the corpus:

@book{korpuspjm2020,
  editor    = {Wójcicka, Joanna and Kuder, Anna and Mostowski, Piotr and Rutkowski, Paweł},
  title     = {Otwarte Repozytorium Korpusu Polskiego Języka Migowego},
  publisher = {Wydział Polonistyki Uniwersytetu Warszawskiego},
  address   = {Warszawa},
  year      = {2020},
  isbn      = {978-83-66400-21-4},
  url       = {https://www.korpuspjm.uw.edu.pl}
}

and our work:

@misc{sienko2026pjm,
  author = {Sieńko, Wojciech and Kovalenko, Rostyslav and Śledzianowski, Albert},
  title  = {A Polish Sign Language to Text Dataset and Benchmark},
  year   = {2026},
  note   = {Polish-Japanese Academy of Information Technology, in preparation}
}

This dataset exists thanks to the Deaf people who recorded themselves for the corpus and the University of Warsaw team who collected and shared it.