MikCil/disambiguation-pretraining-v2

Disambiguation Pretraining v2

Disambiguation Pretraining v2 is an English-Italian corpus for pretraining language models on lexical semantics and word-sense disambiguation. It connects lemmas, definitions, attested contexts, contrasting senses, parts of speech, and subject domains through varied natural-language formulations.

Version 2 contains 3,693,921 rows, compared with 900,009 in the first release. The increase comes from new semantic tasks and controlled reuse of source items across different tasks. Identical normalized training texts are not repeated.

File

• pretraining_text.jsonl: 3,693,921 JSONL records, 2,471,650,411 bytes.
• SHA-256: 3f0762cc1302a5cb70e13ab33e440e325e545546f495746ff76834a31a7d2c8e.

Each record contains:

{
  "id": "pretrain-v2:000000000001",
  "source": "wiktionary",
  "lang": "en",
  "text": "The definition \"...\" describes one sense of \"...\".",
  "metadata": {
    "lemma": "...",
    "sense_uid": "...",
    "task": "gloss_only_single_sense",
    "format": "gloss_only_single_sense",
    "template_id": "GOS-EN-07",
    "semantic_unit_id": "sense:...",
    "render_index": 0,
    "template_seed": 17,
    "build_version": "v2"
  }
}

Metadata varies by task. Contrastive rows may contain multiple sense or instance IDs; domain rows include a normalized domain.

Languages

Language	Rows	Share
English	3,459,119	93.64%
Italian	234,802	6.36%

The imbalance reflects the coverage of the source resources, particularly SemCor, WNGT, and English Wiktionary.

Sources

Source	Rows	Share
Wiktionary	2,927,099	79.24%
SemCor	465,720	12.61%
Mixed-source sense sets	128,254	3.47%
WNGT examples	88,142	2.39%
WordNet inventory	84,706	2.29%

source identifies the source supporting the rendered unit. mixed is used when a contrast or inventory combines senses or examples originating in more than one normalized source.

Tasks

Task	Rows	Share
Single-sense gloss note	964,047	26.10%
Contextual explanation	774,583	20.97%
Definition-to-example link	773,197	20.93%
Same sense in multiple contexts	323,102	8.75%
Domain-conditioned sense	218,044	5.90%
Part-of-speech-conditioned sense	193,307	5.23%
Definition-only contrast	125,030	3.38%
Gloss-only sense inventory	125,029	3.38%
Contextual sense exclusion	54,395	1.47%
Candidate-sense resolution	54,392	1.47%
Pairwise contextual contrast	54,384	1.47%
Three-way contextual contrast	24,619	0.67%
Cross-domain contrast	9,792	0.27%

The build uses 178 distinct template IDs. A semantic item may appear in more than one task, but the same (task, template, semantic unit) is not deliberately repeated and identical normalized output text is removed globally.

Domains

Domains are derived conservatively from retained Wiktionary sense labels through a reviewed exact-match allow-list. Registers, regional labels, temporal labels, and grammatical tags are not interpreted as domains.

The largest English domain counts are biology (33,991), chemistry (29,677), medicine (21,094), computing (10,649), anatomy (8,141), and mathematics (8,134). The largest Italian counts are chemistry (3,527), medicine (3,398), economics (3,275), biology (3,130), zoology (2,254), and history (2,206). Counts include support in domain-conditioned and cross-domain tasks.

Deliberate exclusions

Version 2 does not generate:

• period- or date-conditioned rows;
• register-conditioned rows;
• region-conditioned rows;
• citation-period tasks.

The normalized v1 data does not retain those fields with enough structure and consistency to support reliable claims. They remain candidates for a later build that reparses structured Wiktionary quotations and separates register, region, and temporal metadata.

Quality and validation

The final JSONL was checked in a complete streaming validation pass:

• 3,693,921 sequential records;
• 3,693,921 unique normalized text hashes;
• valid JSON and required metadata on every row;
• no unresolved template placeholders;
• no suspicious mojibake patterns;
• no empty quoted fields;
• no text longer than 1,800 characters;
• no period, register, or citation tasks.

The validation completed with 0 problems. Full counts are in STATS.md, while VALIDATION.md and validation.json record the independent validation results.

Documentation

• IMPLEMENTATION.md: build architecture, filtering, repetition policy, and reproducibility.
• EXPANDED_TEMPLATE_VARIATIONS.md: full candidate template catalog. Only template IDs actually present in pretraining_text.jsonl were used in v2.
• STATS.md and stats.json: descriptive build statistics.
• VALIDATION.md and validation.json: final validation report.

Limitations

• This is pretraining prose, not an SFT question-answer dataset.
• Most Italian data comes from Wiktionary, while the English portion also benefits from SemCor and WNGT gold annotations.
• Wiktionary sense and example quality is heterogeneous.
• Domain extraction is intentionally conservative, but the upstream normalized inventory flattened several Wiktionary label types into one list. The allow-list prevents unreviewed labels from becoming domain claims; a future raw-data reparse could recover more domains.
• WordNet and Wiktionary senses are not force-aligned when an explicit alignment is unavailable.
• Sense frequency in this corpus should not be interpreted as frequency in natural language.