docs/DATA_DOCUMENTATION.md

GLiNER training data documentation

Eladio (eandujar09) created the project's labeled SysML training dataset. He contributed the corpus and preparation code in the September 24, 2026 source contribution. The automatic, rule-based labeling pipeline creates the training annotations from the SysML source texts.

The public dataset and dataset card provide the data, creator credit, preparation details, metadata, exploratory analysis and reuse status. This corpus was used to fine-tune GLiNER RelEx checkpoint 450. Luna is an off-the-shelf model and was not trained on this dataset.

Sources and creation process

The corpus contains 25 accepted SysML documents, principally related linear-actuator examples with additional practice and sequence examples. The original sibling SysML-files directory was not committed, but every accepted document's full text, relative source filename, SHA-256 hash, character annotations and translated SJS are retained in canonical documents.

Eladio's preparation pipeline:

1. Reads source files and removes identical normalized texts using SHA-256.
2. Validates documents through the supplied translator and quarantines rejected inputs. The supplied build records zero errors; acceptance is a syntax/preparation check rather than a guarantee of semantic correctness.
3. Masks embedded @sjs payloads during labeling while preserving character positions.
4. Uses deterministic syntax rules to create entity labels and checks that each character span matches the recorded text.
5. Splits documents using seed 42, then creates non-overlapping chunks of at most 384 source tokens and converts spans to chunk-relative token offsets.
6. Keeps chunks without entities as negative examples and exports the training, validation and test files.

The Hub publication preserves the existing annotations and original split files. It does not add synthetic expansion, augmentation or a new split.

Size and splits

Split	Documents	Context chunks	Labeled entity spans	Entity-free chunks
Train	21	54	1,721	12
Validation	2	5	139	2
Test	2	3	38	0
Total	25	62	1,898	14

The 1,898 figure counts labeled occurrences, not distinct source documents. The Hub's chunks configuration exposes 62 context rows; entities exposes 1,898 occurrence rows from that same corpus. Documents are disjoint across splits. Related design variants occur across splits, so the splits are not independent by design family.

Labels and record format

The 14 labels are package, part definition, port definition, interface definition, requirement, constraint, attribute, action, state, item flow, port, part usage, port definition reference and part definition reference.

Chunk records retain chunk_id, document_id, source_path, source text, tokens, candidate labels and entities. Occurrence records add entity_id, entity text, label, start_token and end_token. Join the configurations with chunk_id. Entity token offsets are inclusive and relative to the chunk; canonical character offsets are half-open. annotation_method is translator-rule.

An actual training occurrence in Linear_Actuator_practice.sysml is PowerOutletPort, labeled port definition, at chunk token offsets 20–20. Source identifiers let readers trace the occurrence to its context. Canonical confidence: 1.0 is a fixed rule marker, not a calibrated prediction.

Exploratory analysis and reproducibility

EDA statistics contain label frequencies, token lengths, source filenames and negative-example counts for each split. The dataset card includes the label distribution plot.

Provenance records Eladio's authorship, source contribution and original-file hashes. SHA256SUMS identifies every published file. From the repository root, run python3 scripts/build_gliner_dataset.py to rebuild the Hub package. This script verifies source hashes, spans, split separation, counts and exact reproduction of the original GLiNER training rows.

Intended use, limits and rights

The data supports reproducible pilot fine-tuning and inspection of SysML entities. Its rules do not supply relationship labels or patent-extraction ground truth. Related examples, class imbalance and narrow source coverage limit generalization. The test split has 38 entities across four labels, so its reported performance does not establish accuracy across all 14 labels or whole patents. Training and evaluation are described in the GLiNER model card.

The separate 100-patent dataset supplies application and comparison inputs; it is not this training corpus. No private user uploads, account credentials or agent conversations were added to the GLiNER release.

Eladio is credited as the dataset creator. The project's training dataset is MIT-licensed: anyone may reuse, modify and redistribute it, including commercially, while retaining the copyright and license notice. See the MIT license and creator attribution. The separate model and patent input collection retain their documented licenses and source rights.