schedula-data / README.md
maxie-12321's picture
dataset build 2026-08-24 09:40
739aeed verified
|
Raw
History Blame Contribute Delete
1.95 kB
---
license: mit
task_categories:
- text2text-generation
tags:
- scheduler
- text-to-json
- tagalog
- taglish
- synthetic
- from-scratch
size_categories:
- 100K<n<1M
---
# Schedula dataset
Fully synthetic, seeded, reproducible corpus for training **Schedula** - a tiny
from-scratch model that turns messy natural language (English / Taglish /
Filipino) into structured schedule JSON.
## Modes
| mode | input | output |
|---|---|---|
| `event_extract` | sentences, bullet dumps, decorated announcements (unicode stars/stylized headers), edit statements; embedded schedule JSON must be ignored | `{"events":[{title, category, date, start_time, duration_min, location, priority}]}` |
| `duplicate_detect` | CURRENT LIST json (+ optional `"time_created": "... ##ignore"` metadata) + numbered NEW ITEMS (1-8) | `{"verdicts":[{duplicate, match_id, confidence, reason}]}` |
## Conventions
- `<today>` anchor date resolves every relative expression ("next week tue",
"bukas", "sa loob ng 2 linggo").
- Fields suffixed ` ##ignore` are app-owned metadata: present in inputs,
never copied to outputs.
- Items under "daily reminders"-style headers are undated (`null`) by design.
- Category labels come from the per-example schema hint (default enum plus
alternate English/Tagalog sets) so the model learns hint-driven mapping.
- Duplicate boundaries include numbered schoolwork (`quiz #2` vs `quiz #3`
never match) and EN<->Filipino translation pairs (do match).
## Splits
{
"train": {
"duplicate_detect": 1282,
"event_extract": 2718
},
"val": {
"duplicate_detect": 184,
"event_extract": 416
},
"test": {
"duplicate_detect": 174,
"event_extract": 426
}
}
Robustness noise (typos/filler/casing) appears only in val/test (16%).
## Reproducibility
Built by `python -m schedula.data.build --config configs/default.yaml`
(seed 1337, created 2026-08-24). The generator lives at
https://github.com/maxie-12321/schedula