NagaYu/ingot-chrono

Ingot — Chrono

Natural-language time expressions to an iCalendar RRULE + ISO-8601 start + IANA timezone

• exception rules, as strict JSON.

Every label in this dataset was constructed before its sentence existed. A schedule object is generated from an integer seed, then rendered into prose. No model, judge or annotator ever decided what the answer was, so the label cannot be wrong -- it is the input to the pipeline.

Splits

split	rows	verified	easy / medium / hard	ja / en
train	48,000	100.00%	10039 / 13920 / 24041	28878 / 19122
val	1,200	100.00%	234 / 344 / 622	737 / 463
test	2,000	100.00%	424 / 553 / 1023	1163 / 837
unseen_combo	2,000	100.00%	0 / 106 / 1894	1241 / 759
unseen_template	2,000	100.00%	493 / 546 / 961	1192 / 808
probe	128	100.00%	29 / 27 / 72	76 / 52

verified is the share of rendered sentences that passed round-trip verification against the constructed label. Rows that failed were discarded, never repaired -- repairing one would mean inferring a label from text, which is the thing this dataset exists to avoid.

Contamination control

Splits occupy disjoint integer seed bands, so a training item and an evaluation item were never produced by the same generator call:

split	seed start	seed stop
train	0	24,000
val	1,000,000,000	1,000,001,200
test	2,000,000,000	2,000,002,000
unseen_combo	3,000,000,000	3,000,002,000
unseen_template	6,000,000,000	6,000,002,000
probe	5,000,000,000	5,000,000,064

unseen_template renders evaluation items using reserved renderer templates that never appear in training. It exists because of a measurement: a blind, hand-written 1,600-line rule parser reaches ~93% exact match on the ordinary test split. That is the ceiling you should expect when the surfaces are themselves rule-generated, so "the specialist beats hand-written rules" is not a claim test can support. Holding surfaces out asks the question the data can answer: does the model generalise to wordings it has never seen? Note the asymmetry -- the rule baseline was developed with all templates visible, so this comparison is biased against the specialist.

unseen_combo is the stronger test of label composition. It draws only from feature combinations that the train/val/test generator is forbidden to emit. Every individual feature value is common in training; only the conjunction is new, which isolates compositional generalisation from memorisation. Reserved combinations (freq, byday_kind, end_kind, holiday_kind):

• DAILY|none|count|after
• DAILY|none|until|skip
• MONTHLY|last|until|after
• MONTHLY|setpos|count|skip
• WEEKLY|multi|until|before
• WEEKLY|single|until|skip
• YEARLY|nth|until|after
• YEARLY|single|count|before

Rendered-sentence collisions with train after generation: val: 2, test: 2, unseen_combo: 0, unseen_template: 0, probe: 0 (rows dropped: 4).

Note that unseen_combo skews hard by construction -- the reserved combinations involve end dates and holiday policies, which score high on the rubric. Compare it against the difficulty- matched slice of test, not against test overall.

Fields

field	meaning
text	the rendered sentence (the model's input)
target_json	the constructed label, canonical compact JSON
reference_date	the "today" relative expressions resolve against
default_tz	workspace default timezone shown in the prompt
difficulty, difficulty_score	derived from the published rubric below
tag_*	structural feature tags
next_occurrences	the next 10 real instants, the oracle for semantic scoring
rrule, dtstart, tzid	the components scored by exact match
verified, verify_mode	round-trip verification outcome

Difficulty rubric (published, so the hard slice is auditable)

axis	value	points
freq	DAILY	0
freq	WEEKLY	0
freq	MONTHLY	1
freq	YEARLY	2
interval_kind	1	0
interval_kind	2	1
interval_kind	3plus	1
byday_kind	none	0
byday_kind	single	0
byday_kind	multi	1
byday_kind	nth	2
byday_kind	last	2
byday_kind	setpos	3
monthday_kind	none	0
monthday_kind	fixed	1
monthday_kind	last	2
end_kind	never	0
end_kind	count	0
end_kind	until	2
dst_kind	none	0
dst_kind	crosses	4
holiday_kind	none	0
holiday_kind	before	3
holiday_kind	after	3
holiday_kind	skip	3
exdate_kind	none	0
exdate_kind	some	2
style_kind	polite	0
style_kind	casual	0
style_kind	terse	0
style_kind	bullet	1
style_kind	email	1
relative_ref	no	0
relative_ref	yes	1

easy = score ≤ 1 · medium = 2–3 · hard = score ≥ 4

Prompt

Prompt version chrono-v1. Training and every evaluation condition share one prompt builder (ingot.prompts), so comparisons measure capability rather than prompt engineering.

Convert the recurring-schedule description into JSON.
Keys, in order: dtstart, tzid, rrule, duration_minutes, exdate, holiday.
- dtstart: local wall time, "YYYY-MM-DDTHH:MM:SS", no timezone suffix. It is the first occurrence on or after today.
- tzid: IANA zone name (e.g. Asia/Tokyo). Use the default unless the text names another.
- rrule: RFC 5545 recurrence rule without the "RRULE:" prefix. UNTIL must be UTC in basic format ending with Z.
- duration_minutes: integer.
- exdate: list of local wall times excluded from the series; [] if none.
- holiday: {"calendar": "JP"|"US"|null, "shift": "none"|"before"|"after"|"skip"}.
Output JSON only.

Reproduce

pip install ingot
python3 scripts/build_dataset.py --n 24000 --variants 2

Generation is a pure function of the seed, so this reproduces the corpus byte for byte.

Limitations

The renderer is rule-based. A sufficiently engineered parser could in principle invert it, so this dataset does not claim that the task is unsolvable by rules. It claims that the labels are strict by construction, that the splits are contamination-free by construction, and that the hard and unseen_combo slices are where general-purpose systems degrade.