phbg/smart-notifications-synthetic

Smart Notifications — synthetic summarization examples

A curated procedural/synthetic dataset for short notification summaries, developed in the collaborative DevTitans project. It is published as a data-preparation contribution, not as evidence of a newly trained model or a benchmark on real users.

Source and authorship

Source: pedrobelota/android-smart-notifications, derived from the collaborative DevTitans repository. Contributor history is preserved there. The source optimizer and its audit report document the transformation from 10,000 source examples to 3,707 unique examples. AI assistants supported portfolio review and publication preparation; no new human annotation campaign is claimed.

The upstream archive/generator version is not independently identified by this export. See provenance.json for the exact curated source commit and file hashes. No additional open-data license has been assigned: public availability is not a blanket grant of reuse rights. Confirm rights with the contributors before redistribution or commercial use.

Data and splits

Split	Rows
train	2,965
validation	371
test	371

Each UTF-8 JSONL row has two string fields: input (notification text) and output (target summary). Targets contain at most 25 whitespace-separated words. Predominantly Portuguese, with multilingual source material; per-row language/category fields were removed in the compact representation.

The audit report records 16 source categories, removal of 6,293 duplicate pairs and zero normalized-input overlap between splits. The prepared files are also checked for exact input duplication, schema and word limits. Exact uniqueness does not imply independence: template families and semantically similar examples can occur in different splits. Split creation is deterministic; it is not a held-out-user or held-out-template design.

Intended use and limitations

Use for SFT pipeline experiments, prompt-format inspection and software tests. No fine-tuning run, model improvement or production-level factuality score is established by publication. This dataset was not used to train the publicly referenced Gemma checkpoint. Several categories contain fewer than 50 distinct pairs; procedural patterns can inflate apparent performance. Evaluate on independently collected, authorized and reviewed data before making quality claims.

Targets favor current facts and corrections over obsolete facts, but automated repair is not equivalent to exhaustive human review. The word-count constraint is a dataset property, not a guarantee that a deployed model obeys it. Notification-like names, amounts, security codes and account scenarios are synthetic examples, not permission to use actual private notifications. Do not use example credentials as real credentials.

Loading

Use the Hugging Face dataset loader with this page's repository ID:

from datasets import load_dataset
# Replace with the repository ID shown on this dataset page.
data = load_dataset("phbg/smart-notifications-synthetic")

Model provenance and Android integration are documented in the source repository's MODEL_CARD.md. Model weights, recordings, tokens, device logs and APKs are not part of this dataset.