| --- |
| license: apache-2.0 |
| language: |
| - en |
| - zh |
| - tl |
| - es |
| - hi |
| tags: |
| - multilingual |
| - translation |
| - world-models |
| - egocentric-video |
| - robotics |
| - annotations |
| - json-schema |
| pretty_name: UniversalLabeler 1.0 |
| size_categories: |
| - n<1K |
| --- |
| |
| # UniversalLabeler |
|
|
| **A loss-audited interchange format for multilingual world-model annotations.** |
|
|
|  |
|
|
| UniversalLabeler separates what happened from how a language describes it. A |
| source annotation is represented as small, evidence-linked claims—action, |
| participants, hand roles, objects, state change, place, time and outcome. Human |
| language captions and dataset-native labels are projections of the same packet, |
| with omissions recorded rather than hidden. |
|
|
| This is a public data and schema release. It does not include a translation |
| service, model prompts or private processing infrastructure. |
|
|
| ## Release 1.0.5 |
|
|
| | | | |
| |---|---:| |
| | Label-language profiles | 19 | |
| | Human languages | 5 | |
| | World-model dimensions | 22 | |
| | Typed annotation kinds | 13 | |
| | Controlled conformance frames | 6 | |
| | Language surfaces | 30 | |
| | License | Apache-2.0 | |
|
|
| Human-language profiles currently cover English, Simplified Chinese, Tagalog, |
| Spanish and experimental Hindi. The controlled examples are test fixtures, not |
| a general-purpose translation benchmark. |
|
|
| ## Data model |
|
|
| ```text |
| native record + evidence |
| │ |
| ▼ |
| ┌─────────────────────────────────────────────┐ |
| │ Universal Annotation Packet │ |
| │ action · roles · hands · objects · state │ |
| │ place · time · outcome · uncertainty │ |
| │ evidence selectors · provenance │ |
| └─────────────────────────────────────────────┘ |
| │ |
| ├── dataset-native label |
| ├── English |
| ├── 简体中文 |
| ├── Tagalog |
| ├── Español |
| └── हिन्दी |
| ``` |
|
|
| The packet is compositional rather than a fixed dictionary of every possible |
| verb and noun. Concepts receive stable identifiers; semantic roles describe |
| their relationship to an event; evidence and source-clock selectors bind each |
| claim to the underlying record. Language-specific grammar remains in the |
| projection layer. |
|
|
| ## Repository contents |
|
|
| | Path | Purpose | |
| |---|---| |
| | `ontology/universal-annotation-core-v1.json` | Node types, predicates and availability dimensions | |
| | `data/label-languages-v1.json` | Dataset and human-language profiles | |
| | `data/world-model-label-audit-v1.json` | Cross-dataset supervision audit | |
| | `data/controlled-action-matrix.jsonl` | Five-language conformance fixtures | |
| | `schemas/` | Strict JSON Schemas for packets, projections and audits | |
| | `examples/fold-towel.uap.json` | Grounded event with hands, state, time and provenance | |
| | `examples/*.adapter.json` | Loss-audited dataset views | |
| | `rubrics/` | Bilingual review and evaluation protocol | |
| | `metadata/` | Reproducibility and validation records | |
|
|
| An interactive explanation is included at `demo/index.html`. Each of its four |
| first-person event images is a generated, synthetic visual preview—not upstream |
| evidence or a source-dataset frame. |
|
|
| ## Recommended workflow |
|
|
| 1. Pin the source dataset, schema revision and native record. |
| 2. Preserve the native record and content digest. |
| 3. Encode only evidence-supported atomic claims. Mark missing dimensions as |
| not collected, not applicable or unknown. |
| 4. Render every target language directly from the same claim packet; do not use |
| one translated language as the source for the next. |
| 5. Record represented and omitted claim IDs for every projection. |
| 6. Validate against the included schemas. |
| 7. Require independent bilingual and evidence-grounded review before accepting |
| generated language as dataset annotation. |
|
|
| ## Evaluation boundary |
|
|
| Release validation checks schema correctness, content addressing, ontology |
| references, semantic equivalence across the controlled matrix, private-data |
| markers and byte-identical rebuilds. Exact results and hashes are in |
| `metadata/validation.json` and `SHA256SUMS`. |
|
|
| These checks establish format conformance. They do not establish open-vocabulary |
| translation quality. Round trips can reproduce the same mistake twice; |
| production releases still require bilingual review against video or other |
| source evidence. The proposed acceptance thresholds and adversarial strata are |
| specified in `rubrics/EVALUATION_PROTOCOL.md`. |
|
|
| ## Known limits |
|
|
| - The six controlled frames cover predicate–patient directives only. |
| - Chinese aspect and classifiers, Tagalog voice/pivot, Spanish morphology, |
| Hindi agreement, code switching, negation and quantifier scope need broader |
| reviewed data. |
| - The supervision audit identifies common world-model dimensions; it does not |
| claim complete adapters for every upstream dataset. |
| - Synthetic receipt identifiers are not evidence of human review. |
|
|
| ## License |
|
|
| UniversalLabeler's original schemas, profiles, ontology and synthetic fixtures |
| are released under Apache-2.0. Upstream datasets, media and annotations retain |
| their own licenses and access conditions and are not redistributed here. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @dataset{universal_labeler_2026, |
| author = {Pablo and contributors}, |
| title = {UniversalLabeler 1.0: Loss-Resistant Translation Contracts for World-Model Labels}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| url = {https://huggingface.co/datasets/itspublu/UniversalLabeler} |
| } |
| ``` |
|
|
