karma689/layer_detection

Tibetan Annotation Layer Detection

266 annotated Classical Tibetan books with seven annotation layers — Quotation, Sabche, Tsawa, Yigchung, Chapter, Author, BookTitle — stored as character-offset spans over a flat base text. Each row is one book: book_id, text, spans ({start, end, label}), and in_scope_layers.

Book-level split: 217 train / 25 validation / 24 test. A book is never split across partitions.

This revision (v1.1) is a span-level clean of v1.0. Texts, book membership, split assignment, and split order are unchanged.

What changed in v1.1

Every span with start == end was dropped, on all seven layers. No other spans were removed. Impossible geometries (start > end, negative offsets, offsets beyond the book text) were scanned and none were found.

Character mass uses inclusive length end - start + 1 (see caveat below).

Layer	Spans before	Spans after	Removed	Char mass before	Char mass after
Quotation	24,730	24,722	8	3,415,105	3,415,097
Sabche	18,847	18,845	2	1,608,291	1,608,289
Tsawa	6,845	6,811	34	1,285,298	1,285,264
Yigchung	9,025	8,804	221	841,535	841,314
Chapter	1,637	1,637	0	94,724	94,724
Author	311	311	0	12,160	12,160
BookTitle	266	266	0	21,938	21,938
Total	61,661	61,396	265	7,279,051	7,278,786

All 265 removals were start == end. 43 books lost at least one span:

Book	Spans removed
P000036	58
P000140	33
P000236	21
P000275	20
P000087	16
P000123	14
P000153	14
P000195	8
P000138	6
P000073	5
P000132	5
P000174	4
P000030	4
P000034	3
P000037	3
P000054	3
P000089	3
P000119	3
P000130	3
P000152	3
P000159	3
P000010	3
P000011	3
P000171	3
P000218	3
P000055	2
P000199	2
P000230	2
P000021	1
P000028	1
P000050	1
P000051	1
P000096	1
P000126	1
P000156	1
P000175	1
P000203	1
P000207	1
P000216	1
P000224	1
P000258	1
P000151	1
P000246	1

v1.0 remains reachable as the v1.0 tag (commit aaa17607).

Why

start == end spans are zero-length or one-character stray marks, not real annotations. The previous validator only checked geometric coherence (negative, inverted, out of bounds), so these passed silently.

Flagged books (zero spans for an in-scope layer)

These books are retained. Their in-scope layer with no remaining spans is unresolved — they are not dropped from the split.

• P000218 (test): 847,725 characters; had 3 Quotation spans, all start == end. After this clean it has no Quotation annotation at all, while still marked in-scope for Quotation.
• P000126 (train): already had zero Quotation spans while marked in-scope. Unchanged by this clean (one non-Quotation start == end span was dropped).

No other in-scope layer on any book went to zero spans as a result of this clean.

Known remaining issues (not fixed here)

• Some books have very low annotation density relative to their length.
• Some books have large positional gaps with no annotation.
• Damage has only been systematically examined for Quotation. The other six layers were cleaned of start == end here but have not been inspected for density or gaps.
• No minimum-length rule, density filter, or gap detection was applied.

The offset convention (inclusive vs half-open) has not been independently verified.

License

Packaging, split, and this clean are released under CC0 1.0. Underlying texts and original annotations come from OpenPecha / Tsadra; per-book source licenses vary.

Config windowed_w8192_s4916 (tag v2.0)

WARNING — derived data

This dataset is derived from a specific tokenizer, window length, and stride. Regenerate it if any of those change. Do not mix windows from different tokenizers or strides in the same training run.

Load the named config, never the unpinned default (that is still book-level):

from datasets import load_dataset
ds = load_dataset("karma689/layer_detection", "windowed_w8192_s4916", revision="v2.0")

revision="v1.1" (default config) stays book texts + character spans. Tag v1.1 is not moved by this publish.

Schedule warning

v1.3 trains on 14,338 windows/epoch (HF stride 5120, step 3072, 62.5% overlap, dataset_revision: null). This config is step 4916 (40% overlap). The continuous extras-scale estimate was **9,060** windows/epoch. The measured HF overflow total is 8,906 (217+25+24 books → 6877+897+1132). That 8,906 is the per-book tokenizer count, not the naive 14338 × 3072/4916 formula and not 9,103 (false 1843-step model).

Relative to 14,338 this is ~38% fewer windows. Warmup and the LR schedule must be rescaled or the same nominal epoch count undertrains.

Provenance

Field	Value
Source dataset	karma689/layer_detection
Source revision	v1.1
Source tag commit	b9576f90b675240a6c59d5d0768afc89735a4bf3
Tokenizer repo (files used)	karma689/mmbert-base-layer-detection-v1.3
Upstream tokenizer	jhu-clsp/mmBERT-base
Vocab size	256000
tokenizer.is_fast	True
Combined tokenizer-file sha256	d582312d8c4ccd73ba5a5199026bd161a869503946e488c5e79deb569150469f
Window length (max_length)	8192
HF tokenizer stride (overlap)	3276
Step (CLI --stride)	4916
Label scheme	15-class BIO softmax
Gold spans split across windows	0
Damaged books (O-mask on loss_mask)	P000218, P000271, P000078, P000126
Measured max span	3095 tokens — P000151 QUOTATION [304112, 308503]

Tokenizer file hashes

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

Hub name s4916 is the step, not the HuggingFace stride argument. tokenizer(..., max_length=8192, stride=3276) yields step 4916.

The longest gold span is 3,095 tokens. A span fits whole in at least one window iff its length is ≤ overlap. Step 4,916 clears that floor with margin.

Window counts

Split	Books	Windows
train	217	6877
validation	25	897
test	24	1132
total	266	8906

Seventeen books are shorter than 8192 tokens; each produced exactly one window. Shortest: P000052, 84 tokens.

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

Offset convention is inclusive [start, end] (text[start:end+1]). Do not convert to half-open — that would change BIO and the v2 token indices.

BookTitle gold stays in labels as B-BOOKTITLE / I-BOOKTITLE. The same positives v1.2/v1.3 ignore are zeroed on loss_mask only.

Overlapped tokens are labelled -100 in every window except the first that covers them, so each content token contributes to the BIO 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

Columns

Column	Description
book_id	OpenPecha id
window_index	0-based window within the book
input_ids / attention_mask	mmBERT window (no padding)
labels	15-class BIO, winner-takes-all (Quotation wins). Unmasked BookTitle.
scope_mask	Per-label in-scope mask
offset_mapping / special_tokens_mask	For char ↔ token maps
spans	Every overlapping gold span (not collapsed): Hub layer, half-open window token start/end, truncated, span_id, char_start/char_end
start_targets / end_targets	[7, T] float, MODEL_LAYERS order, smoothing 1.0/0.3/0.1 with max()
loss_mask	[T] float; 0 on specials/pad, BookTitle BIO tokens, and damaged-book tokens outside every gold span

span_id is {book_id}:{label}:{char_start}:{char_end}:{source_index} and is stable across windows that cover the same gold span.