Datasets:

ampixa
/

nsl-fingerspelling-poses

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
letters: list<item: string>
  child 0, item: string
fluent_vs_beginner_cos: struct<A: double, AA: double, AH: double, AI: double, AM: double, AU: double, BA: double, BHA: doubl (... 484 chars omitted)
  child 0, A: double
  child 1, AA: double
  child 2, AH: double
  child 3, AI: double
  child 4, AM: double
  child 5, AU: double
  child 6, BA: double
  child 7, BHA: double
  child 8, CHA: double
  child 9, CHHA: double
  child 10, DA: double
  child 11, DAA: double
  child 12, DHA: double
  child 13, DHAA: double
  child 14, D_SHA: double
  child 15, E: double
  child 16, GA: double
  child 17, GHA: double
  child 18, GYA: double
  child 19, HA: double
  child 20, I: double
  child 21, II: double
  child 22, JA: double
  child 23, JHA: double
  child 24, KA: double
  child 25, KHA: double
  child 26, KSHA: double
  child 27, LA: double
  child 28, MA: double
  child 29, M_SHA: double
  child 30, NAA: double
  child 31, NGA: double
  child 32, O: double
  child 33, PA: double
  child 34, RA: double
  child 35, TA: double
  child 36, TAA: double
  child 37, THA: double
  child 38, THAA: double
  child 39, TRA: double
  child 40, T_SHA: double
  child 41, U: double
  child 42, UU: double
  child 43, WA: double
  child 44, YA: double
  child 45, YAN: double
rank: struct<A: int64, AA: int64, AH: int64, AI: int64, AM: int64, AU: int64, BA: int64, BHA: int64, CHA:  (... 438 chars omitted)
  child 0, A: int64
  child 1, AA: int64
  child 2, AH: int64
  child 3, AI: int64
  child 4, AM: int64
  child 5, AU: int64
  child 6, BA: int64
  child 7, BHA: int64
  child 8, CHA: int64
  child 9, CHHA: int64
  child 10, DA: int64
  child 11, DAA: int64
  child 12, DHA: int64
  child 13, DHAA: int64
  child 14, D_SHA: int64
  child 15, E: int64
  child 16, GA: int64
  child 17, GHA: int64
  child 18, GYA: int64
  child 19, HA: int64
  child 20, I: int64
  child 21, II: int64
  child 22, JA: int64
  child 23, JHA: int64
  child 24, KA: int64
  child 25, KHA: int64
  child 26, KSHA: int64
  child 27, LA: int64
  child 28, MA: int64
  child 29, M_SHA: int64
  child 30, NAA: int64
  child 31, NGA: int64
  child 32, O: int64
  child 33, PA: int64
  child 34, RA: int64
  child 35, TA: int64
  child 36, TAA: int64
  child 37, THA: int64
  child 38, THAA: int64
  child 39, TRA: int64
  child 40, T_SHA: int64
  child 41, U: int64
  child 42, UU: int64
  child 43, WA: int64
  child 44, YA: int64
  child 45, YAN: int64
clips_used: list<item: string>
  child 0, item: string
files: list<item: struct<path: string, bytes: int64, sha256: string>>
  child 0, item: struct<path: string, bytes: int64, sha256: string>
      child 0, path: string
      child 1, bytes: int64
      child 2, sha256: string
total_bytes: int64
excluded_by_policy: list<item: string>
  child 0, item: string
n_pose_files: int64
to
{'files': List({'path': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')}), 'excluded_by_policy': List(Value('string')), 'n_pose_files': Value('int64'), 'total_bytes': Value('int64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              letters: list<item: string>
                child 0, item: string
              fluent_vs_beginner_cos: struct<A: double, AA: double, AH: double, AI: double, AM: double, AU: double, BA: double, BHA: doubl (... 484 chars omitted)
                child 0, A: double
                child 1, AA: double
                child 2, AH: double
                child 3, AI: double
                child 4, AM: double
                child 5, AU: double
                child 6, BA: double
                child 7, BHA: double
                child 8, CHA: double
                child 9, CHHA: double
                child 10, DA: double
                child 11, DAA: double
                child 12, DHA: double
                child 13, DHAA: double
                child 14, D_SHA: double
                child 15, E: double
                child 16, GA: double
                child 17, GHA: double
                child 18, GYA: double
                child 19, HA: double
                child 20, I: double
                child 21, II: double
                child 22, JA: double
                child 23, JHA: double
                child 24, KA: double
                child 25, KHA: double
                child 26, KSHA: double
                child 27, LA: double
                child 28, MA: double
                child 29, M_SHA: double
                child 30, NAA: double
                child 31, NGA: double
                child 32, O: double
                child 33, PA: double
                child 34, RA: double
                child 35, TA: double
                child 36, TAA: double
                child 37, THA: double
                child 38, THAA: double
                child 39, TRA: double
                child 40, T_SHA: double
                child 41, U: double
                child 42, UU: double
                child 43, WA: double
                child 44, YA: double
                child 45, YAN: double
              rank: struct<A: int64, AA: int64, AH: int64, AI: int64, AM: int64, AU: int64, BA: int64, BHA: int64, CHA:  (... 438 chars omitted)
                child 0, A: int64
                child 1, AA: int64
                child 2, AH: int64
                child 3, AI: int64
                child 4, AM: int64
                child 5, AU: int64
                child 6, BA: int64
                child 7, BHA: int64
                child 8, CHA: int64
                child 9, CHHA: int64
                child 10, DA: int64
                child 11, DAA: int64
                child 12, DHA: int64
                child 13, DHAA: int64
                child 14, D_SHA: int64
                child 15, E: int64
                child 16, GA: int64
                child 17, GHA: int64
                child 18, GYA: int64
                child 19, HA: int64
                child 20, I: int64
                child 21, II: int64
                child 22, JA: int64
                child 23, JHA: int64
                child 24, KA: int64
                child 25, KHA: int64
                child 26, KSHA: int64
                child 27, LA: int64
                child 28, MA: int64
                child 29, M_SHA: int64
                child 30, NAA: int64
                child 31, NGA: int64
                child 32, O: int64
                child 33, PA: int64
                child 34, RA: int64
                child 35, TA: int64
                child 36, TAA: int64
                child 37, THA: int64
                child 38, THAA: int64
                child 39, TRA: int64
                child 40, T_SHA: int64
                child 41, U: int64
                child 42, UU: int64
                child 43, WA: int64
                child 44, YA: int64
                child 45, YAN: int64
              clips_used: list<item: string>
                child 0, item: string
              files: list<item: struct<path: string, bytes: int64, sha256: string>>
                child 0, item: struct<path: string, bytes: int64, sha256: string>
                    child 0, path: string
                    child 1, bytes: int64
                    child 2, sha256: string
              total_bytes: int64
              excluded_by_policy: list<item: string>
                child 0, item: string
              n_pose_files: int64
              to
              {'files': List({'path': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')}), 'excluded_by_policy': List(Value('string')), 'n_pose_files': Value('int64'), 'total_bytes': Value('int64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

NSL Fingerspelling Poses — NSL23-derived

2D whole-body pose sequences for the Nepali Sign Language manual alphabet, extracted from the NSL23 video corpus, plus the letter-endpoint banks and forced-alignment spans derived from them.

This is a derivative of NSL23 (Sunuwar, Borah & Kharga, Sikkim Manipal Institute of Technology), released under CC BY 4.0. Please cite the original dataset as well as this one.

Path	What
`poses/isolated/` (615 `.npz`)	Per-letter clips, one file per NSL23 video
`poses/runthroughs/` (15 `.npz`)	Continuous whole-alphabet takes (fluent signers)
`banks/nsl23_bank_v0.npz`	227 signer×letter endpoint poses, 52 labels (isolated clips)
`banks/fluent_bank_v0.npz`	115 signer×letter endpoints, 46 labels (from run-throughs)
`banks/sweep_5.json`	Forced-alignment letter spans in the run-throughs
`MANIFEST.json`	SHA-256 and byte size of every file

Pose format. Each .npz holds kpts (T, 133, 2) image coordinates, scores (T, 133) confidences, and n (frame count). Layout is COCO-WholeBody 133-keypoint: indices 91–111 left hand, 112–132 right hand. Extracted with RTMPose RTMW-x (rtmw-dw-x-l_simcc-cocktail14_270e-256x192) via rtmlib on CPU. Filenames encode part__conditionFolder__SignerLetter.

Endpoint format. Bank entries are keyed signer|letter and hold a (21, 2) right-hand pose, wrist-centred and palm-scaled (divided by the wrist→middle-MCP distance), taken from the most stable high-confidence window of the clip.

What this is useful for

Signer-independent handshape recognition, endpoint banks for fingerspelling synthesis, and pose-space transfer experiments. All 630 NSL23 videos extracted with zero pose failures.

Honest limitations

Read these before using the data.

The isolated clips are largely a beginner corpus. NSL23's five fluent signers (S3–S6, S14) recorded the alphabet as continuous single-shot run-throughs, not per-letter clips. Isolated per-letter clips come mostly from beginners who learned the handshapes from a chart immediately before recording.
The forced-alignment spans in sweep_5.json have unreliable boundaries. Span identity passes a shuffle control (permuted letter order, tau chosen on consonants, held-out vowels median z=3.2). Span boundaries lose to an even-split null when checked against human-cut clips (1.2–3.9 s error vs 0.8–1.0 s). Endpoints survive this because the stable-hold search finds the right moment inside a sloppy span; inter-letter transitions derived from these spans should not be trusted.
Every NSL23 take has a cropped low-resolution duplicate. Near-duplicates must never straddle a train/test split.
Label hygiene. NSL23 filenames contain inconsistencies (D_SA for D_SHA, KSH and K_SHA for KSHA, DHHA); unambiguous cases are normalised and the rest quarantined. Nomenclature follows the NSL23 documentation: TA=ट (retroflex), TAA=त (dental), T_SHA/M_SHA/D_SHA = श/ष/स.
All gestures are right-handed. Signer coverage is uneven (S1/S2 appear in eight condition folders each; S7–S13 in one).

Deliberately excluded

This release contains only NSL23-derived material. It excludes, by policy:

Anything derived from the NDFN / deafnepal.org.np dictionary (rights-reserved, not for redistribution).
Parliament broadcast poses and weakly-aligned pairs.
Any material from consented human recording sessions, whose participants were promised their recordings would not be publicly distributed.

Citation

Cite NSL23 first:

@article{sunuwar2024nsl23,
  title   = {NSL23 dataset for alphabets of Nepali sign language},
  author  = {Sunuwar, Jhuma and Borah, Samarjeet and Kharga, Aditi},
  journal = {Data in Brief},
  year    = {2024},
  doi     = {10.1016/j.dib.2024.110158}
}

Then this derivative:

@misc{ampixa2026nslfingerspell,
  title  = {NSL Fingerspelling Poses (NSL23-derived)},
  author = {Ampixa},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ampixa/nsl-fingerspelling-poses}
}

Provenance

Extraction and analysis code: src/fingerspell_bank.py, src/segment_runthroughs.py, src/build_fluent_bank.py, src/nsl23_bank_analysis.py in the (private) Ampixa NSL research repository. Method notes and every result quoted above are in docs/FINGERSPELL.md.

Downloads last month: 25

Total file size:

143 MB