fosters/homograph-bel-contextual-v1

HomographBel Contextual v1

Prepared contexts for fine-tuning and evaluating contextual Belarusian homograph stress resolution. Each row describes one exact target occurrence and preserves its source text, half-open character span, dictionary identifiers, provenance, quality tier, orthography, grouped split, and sampling weight.

This is the training dataset for fosters/homograph-bel-xlm-roberta-base.

Dataset summary

• Dataset ID: a747519e553a3532075e0f30517175ee93cdc93ce6ed448ab1ac7e912521b412
• Accepted contexts: 107,976
• Accepted homographs: 1,942
• Observed variants: 2,898
• Orthography: official_2008
• Dictionary: GrammarDB-RELEASE-202601
• Preparation schema: version 1

Split	Rows	Intended use
train	102,029	Optimization only
silver_validation	5,134	Broad model-labelled diagnostics
gold_validation	275	Human-reviewed model selection
gold_test	538	Reserved human-reviewed final evaluation

Training label tiers are 2,128 gold, 18,390 silver-high, and 81,511 silver-medium. The complete dataset has 2,941 gold, 19,333 silver-high, and 85,702 silver-medium examples.

Load

from datasets import load_dataset

dataset = load_dataset("fosters/homograph-bel-contextual-v1", token=True)
print(dataset["train"][0])

To reproduce the exact local artifact rather than the viewer-ready split shards:

from huggingface_hub import snapshot_download
from pathlib import Path

root = Path(snapshot_download(
    "fosters/homograph-bel-contextual-v1",
    repo_type="dataset",
))
examples = root / "metadata" / "examples.parquet"
dictionary = root / "metadata" / "dictionary.parquet"

The metadata/ directory also contains the immutable preparation manifest and quality report. The review queue and rejected examples are intentionally excluded because they are not training data.

Important fields

• text: unchanged source context
• target_start, target_end: exact half-open Python character span
• target_surface: text[target_start:target_end]
• canonical_official_stressed: dictionary-approved target with stress
• homograph_id, variant_id: stable dictionary identifiers
• orthography: orthographic system; currently official_2008
• annotation_evidence: source and annotator provenance
• quality_tier: gold, silver_high, or silver_medium
• source_group_id: grouping key used to prevent source leakage across splits
• split: fixed prepared split; do not randomly resplit these rows
• sampling_weight: example-level training weight before quality-tier multiplication

Sources and annotation

Contexts come from Belarusian Common Voice sentence material and the Belarusian subset of FineWeb2. Stress labels combine human review and Gemini-family model annotation. The preparation pipeline normalizes labels against GrammarDB, preserves provenance, deduplicates accepted contexts, and keeps sentence/source groups within one split.

FineWeb2 Round 1 and Round 2 are unfiltered web contexts. Potentially toxic, sensitive, or identifying content can remain. Content flags are provided for selected patterns but are not a complete safety classifier.

Limitations

• Only 2,941 of 107,976 accepted contexts are human-reviewed gold labels.
• Gold validation is small: 275 contexts, 28 homographs, and 43 observed variants. Only 15 homographs expose both variants in validation.
• Silver validation is useful for coverage diagnostics but not for an unbiased quality claim.
• Homograph coverage is highly imbalanced; many homographs expose only one observed variant.
• The complete GrammarDB dictionary contains entries without accepted training evidence.

License and access

This preliminary repository is private. FineWeb2 declares ODC-By, while the Common Voice material and the derived annotations have separate provenance and attribution considerations. Because the combined redistribution terms have not yet been signed off, this dataset is marked license: other and should not be made public until a source-by-source licensing and content review is complete.

Reproducibility

The canonical artifact hashes, input versions, preparation configuration, and logical content hashes are recorded in metadata/manifest.json. Always identify experiments by dataset ID and repository revision, not by the repository name alone.