Datasets:

karma689
/

layer_detection_windowed_w8192_s4916

input_ids listlengths 84 8.19k	attention_mask listlengths 84 8.19k	special_tokens_mask listlengths 84 8.19k	offset_mapping listlengths 84 8.19k	labels listlengths 84 8.19k	scope_mask listlengths 15 15	book_id stringclasses 217 values	window_index int64 0 132
[2,235248,244146,246634,238324,109339,237831,238274,35773,238679,238277,122663,238134,245916,238034,(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[0,1],[0,1],[1,2],[2,3],[3,5],[5,6],[6,7],[7,9],[9,10],[10,11],[11,14],[14,15],[15,16],[16,17(...TRUNCATED)	[-100,13,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,14,1(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	0
[2,238277,35773,185434,237819,103158,241027,210270,238679,242386,55610,238808,35773,241027,111088,23(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[7427,7428],[7428,7430],[7430,7433],[7433,7434],[7434,7436],[7436,7437],[7437,7440],[7440,744(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	1
[2,238808,35773,237304,241796,238034,152582,237819,238808,190316,240801,238458,55610,238808,109684,2(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[14885,14886],[14886,14888],[14888,14889],[14889,14890],[14890,14891],[14891,14893],[14893,14(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	2
[2,93673,238134,242386,238034,80716,164570,93673,241027,96625,237304,243472,238627,55610,237304,2439(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[22350,22352],[22352,22353],[22353,22354],[22354,22355],[22355,22357],[22357,22360],[22360,22(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	3
[2,242837,238277,55610,103158,238451,241211,237780,49513,80716,241027,190316,237831,49513,185434,238(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[29753,29754],[29754,29755],[29755,29757],[29757,29759],[29759,29760],[29760,29761],[29761,29(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	4
[2,49513,237831,48790,238134,243259,65616,96625,237937,241796,236215,103158,238451,241850,176554,118(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[37353,37355],[37355,37356],[37356,37358],[37358,37359],[37359,37360],[37360,37362],[37362,37(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	5
[2,241501,238458,206087,109339,240801,238277,65616,109684,238134,152582,238451,219468,237304,240789,(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[44770,44771],[44771,44772],[44772,44774],[44774,44776],[44776,44777],[44777,44778],[44778,44(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	6
[2,244296,103158,238808,206087,109339,238134,152582,238451,241211,238547,65616,241317,111066,240268,(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[52295,52296],[52296,52298],[52298,52299],[52299,52301],[52301,52303],[52303,52304],[52304,52(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	7
[2,240268,49513,237831,241198,48790,238451,238277,80716,241317,55610,185434,238547,237831,239791,238(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[59857,59858],[59858,59860],[59860,59861],[59861,59862],[59862,59864],[59864,59865],[59865,59(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	8
[2,237304,240980,55610,238808,220846,238627,65616,96625,237819,238277,35773,238679,65616,152582,2382(...TRUNCATED)	[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1(...TRUNCATED)	[1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0(...TRUNCATED)	[[0,0],[67402,67403],[67403,67404],[67404,67406],[67406,67407],[67407,67410],[67410,67411],[67411,67(...TRUNCATED)	[-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100,-100(...TRUNCATED)	[ 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 ]	P000012	9

End of preview. Expand in Data Studio

Tibetan Layer Detection — windowed (w8192, step 4916)

Derived data. This is not a new annotation set. It is karma689/layer_detection v1.1 tokenized with a specific mmBERT tokenizer into 8,192-token sliding windows. Regenerate it if the tokenizer, window length, or stride changes. Do not mix windows from different tokenizers or strides in the same training run.

Each row is one window over a Classical Tibetan book, with BIO labels for seven annotation layers (Quotation, Sabche, Tsawa, Yigchung, Chapter, Author, BookTitle). Books never cross split boundaries.

How to load

from datasets import load_dataset

ds = load_dataset("karma689/layer_detection_windowed_w8192_s4916")
print(ds)
# DatasetDict({
#     train: Dataset({ features: [...], num_rows: 6877 }),
#     validation: Dataset({ features: [...], num_rows: 897 }),
#     test: Dataset({ features: [...], num_rows: 1132 }),
# })

Provenance

Field	Value
Source dataset	`karma689/layer_detection`
Source revision	`v1.1` (the `start == end` clean)
Source tag commit	`b9576f90b675240a6c59d5d0768afc89735a4bf3`
Source data commit	`ac507313bce467e609d83296a3bcd9243e08a170`
Tokenizer files loaded from	`karma689/mmbert-base-layer-detection-v1.3`
Upstream tokenizer (v1.3 training config)	`jhu-clsp/mmBERT-base`
Vocab size	256,000
`tokenizer.is_fast`	`True`
Combined tokenizer-file sha256	`d582312d8c4ccd73ba5a5199026bd161a869503946e488c5e79deb569150469f`
Window length (`max_length`)	8,192
HF tokenizer `stride` (overlap)	3,276 tokens (39.99%)
Step (`max_length − stride`)	4,916 (the `s4916` in the repo name)
Training configs (v1 / v1.2 / v1.3)	yaml `stride: 5120` = step 3,072, 62.5% overlap — read from `experiments/mmbert_base_v1.3/config.yaml`, not assumed
Gold spans split across windows	0
Measured max span	3,095 tokens — P000151 `QUOTATION` `[304112, 308503]` (validation)
On-disk size before push	2.300 GiB

The v1.3 checkpoint does ship tokenizer.json / tokenizer_config.json; those files were used. They are not byte-identical to the upstream jhu-clsp/mmBERT-base snapshot, but they share vocab size 256,000 and produced matching encodings on a sample.

Tokenizer file hashes

File	sha256
`tokenizer.json`	`609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f`
`tokenizer_config.json`	`14b147f2a4f939d9b12ab36e9633917040dd948fa78ce283b03402e4cf2c9cba`

Why this window / stride

HuggingFace's tokenizer stride is overlap, not step. The call that built this set is:

tokenizer(text, truncation=True, max_length=8192, stride=3276,
          return_overflowing_tokens=True, return_offsets_mapping=True)

That yields step 8192 − 3276 = 4916. The Hub name uses the step.

A gold span appears whole in at least one window iff its length is ≤ overlap. The longest span in this corpus is 3,095 tokens, so the zero-split floor is step ≤ 8192 − 3095 = 5097. Step 4,916 clears that with margin. Do not increase the step past that floor — that starts cutting real spans.

Measured overlap on consecutive full-size windows: 3,276 tokens on 8,632 pairs, 3,277 on 8 pairs (HF overflowing at a token boundary). Never below 3,276.

Window counts

Split	Books	Windows
train	217	6,877
validation	25	897
test	24	1,132
total	266	8,906

Book membership and split assignment are identical to source v1.1. Seventeen books are shorter than one full window; each produced exactly one window and was not dropped. Shortest: P000052, 84 tokens.

Source spans are inclusive [start, end] (text[start:end+1]), matching layer_detection.training_utils.align_window_labels.

Schema

Each row is one window:

Column	Type	Description
`book_id`	string	OpenPecha book id (e.g. `P000012`)
`window_index`	int	0-based window within that book
`input_ids`	list[int]	Token ids, length ≤ 8,192, not padded
`attention_mask`	list[int]	1 on every position (no padding)
`labels`	list[int]	15-way BIO ids; `-100` on specials, empty offsets, and stride-prefix tokens already supervised in an earlier window
`scope_mask`	list[int]	Length-15 mask: which label ids this book's `in_scope_layers` may contribute to the softmax/loss
`offset_mapping`	list[[start, end]]	Half-open character offsets per token
`special_tokens_mask`	list[int]	1 on CLS/SEP (and other specials)

scope_mask uses the same confirmed-scope policy as v1.3 training (apply_confirmed_scope): P000010 excludes Quotation, P000100 excludes Tsawa.

BookTitle gold is not replaced with -100 here. Training-time masked_layers remains a training concern.

Overlapped tokens are labelled -100 in every window except the first that covers them (already_covered_through in align_window_labels), so each content token contributes to the loss once.

15-label id → name (v1.3)

id	label
0	`O`
1	`B-QUOTE`
2	`I-QUOTE`
3	`B-SABCHE`
4	`I-SABCHE`
5	`B-TSAWA`
6	`I-TSAWA`
7	`B-YIGCHUNG`
8	`I-YIGCHUNG`
9	`B-CHAPTER`
10	`I-CHAPTER`
11	`B-AUTHOR`
12	`I-AUTHOR`
13	`B-BOOKTITLE`
14	`I-BOOKTITLE`

Citation

Cite this derived set, the source book-level dataset, and OpenPecha / Tsadra as the origin of the texts and annotations:

@misc{tibetan_layer_detection_windowed_w8192_s4916,
  title  = {Tibetan Annotation Layer Detection Dataset (windowed, w8192 s4916)},
  author = {karma689},
  year   = {2026},
  url    = {https://huggingface.co/datasets/karma689/layer_detection_windowed_w8192_s4916},
  note   = {Derived from karma689/layer_detection v1.1; tokenizer karma689/mmbert-base-layer-detection-v1.3; source texts and annotations from OpenPecha / Tsadra Foundation}
}

@misc{tibetan_layer_detection,
  title  = {Tibetan Annotation Layer Detection Dataset},
  author = {karma689},
  year   = {2026},
  url    = {https://huggingface.co/datasets/karma689/layer_detection},
  note   = {Source texts and annotations from OpenPecha, built on editions curated with Tsadra Foundation}
}

License

Packaging of this windowed view is released under CC0 1.0 Universal, same as the source dataset's packaging, split, and span clean.

Underlying Tibetan texts and original annotations come from individual OpenPecha-Data repositories; per-book source licenses vary. Check those before redistribution beyond research/training.

Acknowledgements

Texts and original layer annotations were produced through OpenPecha from editions curated with Tsadra Foundation. This repository only tokenizes and windows the already-cleaned v1.1 book-level dataset.

Downloads last month: -