CrocoSign/pjm-segments

PJM sentence clips (unfiltered)

All 39,817 sentence clips cut from the Open Repository of the PJM Corpus (Sign Language Linguistics Lab, University of Warsaw), each with its Polish transcript, as WebDataset shards. This is the step before selection: the ready-to-train dataset of 29,928 pairs, with English translations, splits and features, is CrocoSign/pjm-data.

Format

89 tar shards, train-000000.tar to train-000088.tar, about 25 GB in all. Each clip is two members:

• {key}.mp4: the video of one sentence, one signer, H.264. Despite the extension, the container is MPEG-TS; browsers need it remuxed to MP4 (ffmpeg -i clip.mp4 -c copy -movflags +faststart out.mp4).
• {key}.txt: the Polish transcript of that sentence, UTF-8, as cut from the corpus subtitles (the corrected texts of the 29,928 dataset clips are in pjm-data/texts/texts_fixed.csv).

The key, such as 004_000_M, is {recording}_{sentence}_{side}, where the side is L or R for a half of a dialogue recording and M for a monologue. It is the same key as in pjm-data, so it joins the clips with the splits, features and poses there.

from datasets import Value, load_dataset

clips = load_dataset("CrocoSign/pjm-segments", split="train", streaming=True)
clips = clips.cast_column("mp4", Value("binary"))  # raw bytes; decoding video would need torchcodec
first = next(iter(clips))
print(first["__key__"], first["txt"])  # 004_000_M Wiesz, kot i ptak w klatce.
video = first["mp4"]["bytes"]

Streaming reads the shards in order, so finding one clip by key this way scans the dataset. To fetch a single clip, request its byte range from the shard instead. clip_index.csv gives, for every clip, the shard and the offset and size of its .mp4 member:

# 682_073_R,train-000085.tar,352468480,471880
curl -L -r 352468480-352940359 -o 682_073_R.ts \
  https://huggingface.co/datasets/CrocoSign/pjm-segments/resolve/main/train-000085.tar

What was left out of pjm-data, and why

Reason	Clips
one-word sentence	5,527
fewer than three words (final quality filter)	3,603
only very common words (final quality filter)	362
dominated by rare words (seen at most three times)	362
longer than 60 sub-word tokens	19
translation flagged as invalid	10
broken video (one to five frames)	4
empty sentence	2

Licence and attribution

CC BY-SA 4.0, as the corpus. The clips show the faces of the signers who recorded themselves for the corpus. Cite the corpus (Wójcicka, Kuder, Mostowski, Rutkowski (eds.), Otwarte Repozytorium Korpusu Polskiego Języka Migowego, University of Warsaw, 2020, ISBN 978-83-66400-21-4) and our work (Sieńko, Kovalenko, Śledzianowski, A Polish Sign Language to Text Dataset and Benchmark, 2026, in preparation); BibTeX for both is in the pjm-data card.