VoxSign/USL-Gloss

VoxSign USL Gloss Dataset

Authors: Taremwa Aaron Francis · Tusuubira Victor

English → Ugandan Sign Language (USL) gloss approximation pairs, for fine-tuning sequence-to-sequence models.

⚠ The gloss convention changed on 2026-09-05. Re-download if you pulled this before then.

A native USL signer reviewed the convention and five rulings were adopted. Roughly 79% of the targets in this corpus changed. If you have an older copy, it encodes a superseded convention — most visibly it strips indefinite articles, hyphenates compounds, drops noun plurals and forces verbs to base form. See What changed.

The previous build remains reachable by commit SHA if you need to reproduce earlier work.

Read this before using the dataset

The gloss side is rule-generated, not written by signers, and flat uppercase gloss is not USL — it cannot carry spatial agreement, classifier predicates, non-manual grammatical marking or verb directionality. One signer has now reviewed one axis of the convention; word order, function-word retention and lexicon coverage remain unadjudicated inferences. Treat this as a consistent gloss approximation for bootstrapping, not as ground truth for Ugandan Sign Language.

Splits

Split	Rows	Origin
train	23,620	Tatoeba English, glossed by the VoxSign rule engine
validation	5,905	20% split of the same pool
test	308	Hand-authored gold set — never machine-generated

The test split is hand-authored deliberately. A test set produced by the same engine that produced the training targets measures only how well a model imitates that engine. It shares zero sources and zero targets with train or validation, verified by an automated gate on every build.

The test split is gold v2: the same 308 human-written targets as before, with only the morphology the signer review changed. Every human judgement about word order, clause segmentation and word choice is untouched, and the rule engine was never run over it — a reference generated by the system under test measures nothing.

Format

{
  "input_text": "translate English to USL gloss: my mother is coming to beat me",
  "target_text": "MY MOTHER COMING TO BEAT ME"
}

Gloss is ALL CAPS, unpunctuated, with clauses in a long utterance separated by //.

from datasets import load_dataset
ds = load_dataset("VoxSign/USL-Gloss")

The translate English to USL gloss: prefix is part of the stored input, not something to add at training time.

Composition

	Value
Distinct English words	18,772
Distinct gloss tokens	18,540
Input length range	3 – 42 words
Multi-clause rows	33.5% (9,881)
Noise-augmented inputs	20%

Length distribution, deliberately spread so a model cannot learn a single sentence shape:

words	rows	share
3–8	8,778	29.7%
9–15	10,130	34.3%
16–25	7,089	24.0%
26+	3,508	11.9%

Noise augmentation. 20% of training inputs are perturbed copies — lowercased, punctuation stripped, apostrophe-less contractions (dont, im, theres), disfluencies — while the target stays the gloss of the clean sentence. Without this a model would only learn to imitate the rule engine on tidy input. Real input arrives from typing and speech recognition, which are neither.

How it was built

Tatoeba English ─▶ quality + safety filter ─▶ length stratification ─▶ rule engine ─▶ gates
                                                                            │
                                            + salvaged v1 rows ─────────────┤
                                            + synthesised multi-clause ─────┘

1. Source. English sentences from Tatoeba (CC-BY 2.0 FR) — a conversational register, which matters more than raw size for this task.
2. Filtering. Rows are rejected for malformed text, missing verbs, templated shapes, and content safety. Tatoeba is crowd-sourced and contains slurs and bigoted quotations; none of that belongs in training data for an accessibility tool.
3. Long tail. Tatoeba thins out above 20 words, so longer items are built by joining two human-written sentences with a real connective. This is the only machine-generated English, and it never slot-fills.
4. Glossing. A rule engine over POS-tagged English (NLTK perceptron tagger + WordNet) applies the published specification.
5. Gates. The build fails on any templated row, any train/test leakage, any duplicate, any spec violation, vocabulary under 8,000 words, or a length band under 5%.

Current build, every gate passing:

Gate	Result
Templated rows	0
Test source or target leaked into train	0 of 308
Duplicate inputs	0
Targets violating the spec	0
English vocabulary ≥ 8,000	18,772
Length bands ≥ 5% each	4 / 4

What changed in the 2026-09-05 revision

The convention was reviewed by Sandra Naikambo, a native USL signer, against the cases where the rule engine and the salvaged hand-written gloss disagreed (505 rows once machine-generated template rows were excluded; the engine already agreed with 403 of them, 79.8%). Her reading outranks the specification's earlier inferences, which came from salvaged data and general sign-language practice rather than observation.

Ruling	Before	After
Keep indefinite articles	ME NEED DOCTOR	ME NEED A DOCTOR
Compounds as two words	PRIMARY-SCHOOL, LAST-YEAR	PRIMARY SCHOOL, LAST YEAR
Keep noun plurals	MY PARENT	MY PARENTS
Keep verb inflection	MY MOTHER COME	MY MOTHER COMING
Mark expletive present	CLOUDY	CLOUDY NOW

The article ruling covers the indefinite article only. All seven rows behind it are a or an — I need a doctor, take a picture. None involves the, the reviewer was never asked about the definite article, and so THE is still dropped. That is not a reversal of her ruling; it is declining to extend it on no evidence, where doing so produced TODAY THE BEST DAY OF MY LIFE. It is also what sign linguistics predicts: definiteness is carried spatially by pointing at an established locus, so the has no manual sign, while indefinite a is meaningful and often maps to ONE.

Two lines in the spec are the maintainers' decisions, not the signer's, and are marked as such: RAINING NOW (she wrote RAIN NOW) and INSTALL THE APPLICATION (she wrote INSTALL APPLICATION). Both resolve a conflict in favour of the general rule over a single row. Word order was not ruled on — the reviewer correctly abstained, because most rows in that group were legacy template output rather than anyone's judgement.

Full detail is in §9b of USL_GLOSS_SPEC.md, published in this repository.

Why this corpus was rebuilt

The v1 corpus was 79% machine-generated from templates producing ungrammatical, semantically impossible English — "I studys at the bank", "They does not travel food", "Tomorrow i will cook computer". Its test split was 35% the same templates, so evaluation measured template memorisation and looked healthy while the deployed model returned SHE MY BIRTHDAY for "she is my girlfriend".

	v1	current
Templated rows (train)	4,742 / 6,000 (79%)	0
Test rows from the training generator	524 / 1,500 (35%)	0
Distinct English words	596	18,772
Longest input	8 words	42 words
Multi-clause coverage	~0%	33.5%

Of the original 7,500 rows, 6,970 were discarded as machine-templated; 530 were salvaged and re-glossed to the current specification. Those 530 are the only human-authored gloss in the project, and the rule engine currently agrees with 76.0% of them.

Gloss convention

Published as USL_GLOSS_SPEC.md in this repository, and binding on the corpus builder, the training targets and the production output validator alike.

[TIME] [SUBJECT] [VERB] [OBJECT / COMPLEMENT] [MODAL] [NEG] [WH]

I am going to the hospital   ->  ME GO TO HOSPITAL
Where is the bathroom?       ->  BATHROOM WHERE
I do not like beans          ->  ME LIKE BEANS NOT
Can you help me?             ->  YOU HELP ME CAN
I need a doctor              ->  ME NEED A DOCTOR
I am tired, I am going home  ->  ME TIRED // ME GO HOME

Limitations and bias

• Not USL. See the notice at the top. One signer has reviewed morphology and article retention. Word order, whether the retained function words correspond to real signs, whether past tense should be marked, and whether modal-final order holds across all modals are all still unadjudicated.
• Rule-generated targets. The gloss side inherits the rule engine's assumptions and its POS tagger's errors. Spot-check before trusting any individual row.
• Only 530 human-authored rows exist, and they informed the rules rather than appearing as targets in bulk. More human-authored gloss is the single thing that would most improve this corpus.
• Source-domain bias. Tatoeba skews toward everyday European and North American conversational content. Ugandan domain vocabulary is seeded (matooke, boda-boda, Kampala, Luganda, mobile money) but shallow.
• Safety filtering is coarse. It is deliberately over-inclusive and will not be exhaustive.
• English-only source. No Luganda, Swahili or other Ugandan language input.
• 34 rows of the gold test split are unmeasured on one dimension: where an indefinite article belongs inside a reordered gloss is a placement judgement, so those rows were deliberately not auto-generated. They need a person.

Before output trained on this data reaches Deaf users, the specification and a sample of the gold set should be reviewed further by native USL signers, ideally through UNAD (Uganda National Association of the Deaf). The lexicon is editable JSON and the spec is a single document specifically so that review can be applied without hand-editing a corpus.

Models trained on this dataset

VoxSign/Speech-to-Gloss — FLAN-T5-Base fine-tuned on this corpus. Note that VoxSign serves gloss with the rule engine by default, not the model: on the gold test set the engine leads it by 3.8 BLEU, and the ceiling is structural, because the engine generated these training targets in the first place.

Licence and attribution

CC-BY 2.0 FR, inherited from the source text.

English sentences from the Tatoeba Project, released under CC-BY 2.0 FR. Gloss targets generated by the VoxSign rule engine.

Convention review: Sandra Naikambo (native USL signer), 2026-09-05.

Citation

@misc{voxsign_usl_gloss_dataset,
  title  = {VoxSign USL Gloss Dataset},
  author = {Taremwa, Aaron Francis and Tusuubira, Victor},
  year   = {2026},
  note   = {English from Tatoeba (CC-BY 2.0 FR); gloss targets rule-generated,
            convention reviewed by a native USL signer 2026-09-05.
            A gloss approximation, not Ugandan Sign Language.},
  url    = {https://huggingface.co/datasets/VoxSign/USL-Gloss}
}