NoahIsARider/temporaledge

TempEdge

TempEdge is a bilingual (English / Chinese) benchmark of 10,900 machine-verified questions about local time, DST transitions, calendar arithmetic, and business calendars. Every label is produced by an executable reference implementation and re-derived by an independent verifier; there is no human annotation and no model-generated label.

Why this dataset exists

Public temporal-reasoning data is dominated by contextual time question answering (TimeQA, TempReason, MenatQA) and by city-level time difference probes (GeoTemp). Those benchmarks test reading comprehension and world knowledge. They rarely isolate the failure mode that shows up in production tool use: converting a wall-clock time that is skipped or repeated by a DST transition, adding elapsed hours versus calendar days, or landing on the right business day across a holiday calendar.

Field reports keep pointing at the same root cause: LLMs and date libraries get the rules right in isolation and fail when several rules interact, especially around DST edges and historical offset changes.

TempEdge fills that gap with tasks whose answers are exact, short, and mechanically checkable, so evaluation needs no LLM judge and no fuzzy date parsing.

Task families

family	what it asks	answer type	n
wall_to_utc	local wall clock in an IANA zone → UTC, including times that are skipped or repeated	UTC datetime or nonexistent	1200
utc_to_wall	UTC instant → local datetime plus numeric offset	local datetime with offset	1000
add_absolute	add elapsed hours across a transition, then render locally	local datetime with offset	900
add_wall	add calendar days of wall-clock time, then render locally	local datetime with offset	900
business_add	move N business days using official public holidays	date	900
business_count	count business days in a half-open interval	integer	900
iso_week	ISO 8601 week-year and week number, including the 53-week years	YYYY-Www	600
month_clamp	add calendar months with month-end clamping	date	700
weekday	weekday name for a date in 1900–2100	weekday name	600
epoch_to_local	Unix timestamp → local datetime with offset	local datetime with offset	800
offset_at	UTC offset of a zone at an exact instant, including historical offsets	±HH:MM[:SS]	900
next_transition	exact UTC instant of the next offset transition	UTC datetime	700
classify_local	decide whether a local time is normal, ambiguous, or nonexistent	enum	800

Difficulty mix: 1,840 easy / 3,447 medium / 5,613 hard. Difficulty is assigned per task family rather than per row, so the mix above describes how many rows sit in the five families that are hard by construction (DST gaps and folds, 30-minute DST zones, historical offsets, holiday calendars with non-Saturday weekends). Within wall_to_utc, offset_at, utc_to_wall and epoch_to_local a row is easy when the zone is fixed and hard when a transition is involved. Read the per-difficulty table as a family breakdown, not as a calibrated item difficulty scale.

Fields

field	description
id	stable row id
family	task family
difficulty	easy / medium / hard
inputs	machine-readable inputs (timezone, local datetime, UTC instant, country, etc.)
question_en	English prompt, self-contained and answer-format explicit
question_zh	Chinese prompt, semantically equivalent
answer	canonical answer string
answer_type	schema used for validation
tzdata_version	IANA tzdata release used for labels
holidays_version	holidays package release used for business calendars
generator_version	generator release

Splits: dev (545 rows, stratified per family, for prompt tuning) and test (10,355 rows).

Provenance and verification

• Timezone labels: Python zoneinfo backed by IANA tzdata 2026.5, including historical and future rules.
• Business calendars: holidays 0.107 with each country's own weekend definition (for example Israel uses Friday–Saturday).
• Generation seed: 20261020, generator 1.0.1. scripts/generate.py reproduces the data byte-for-byte when the three versions above are pinned, because the labels are functions of the tz database. A tzdb release that amends a historical rule moves the affected labels, which is why every row carries its own tzdata_version.
• scripts/verify.py re-derives every label through an independent code path that does not use fold. It enumerates candidate UTC instants for a local time and it locates transitions by scanning offsets rather than reusing the generator cache.

Verification result for this release, re-run with tzdata==2026.5 and holidays==0.107: 10,900 / 10,900 rows pass, zero duplicate ids, zero format violations. The 1.0.1 release corrects one wall_to_utc row on America/Bogota. The 1992 Colombia transition sits at 1992-05-02 00:00 local in tzdata 2026.5, so the shifted wall time in the earlier build had drifted one day past the gap.

Evaluation

Answers are short strings, so exact match after light normalization is the metric:

• trim whitespace, case-insensitive for weekday and enum answers
• accept 2024-01-01T00:00:00+00:00 for 2024-01-01T00:00:00Z
• accept a space instead of T in ISO datetimes

python scripts/eval.py --pred predictions.jsonl --split test

predictions.jsonl is one JSON object per line with id and prediction (CSV with id,prediction also works). The script reports exact match overall, per family, and per difficulty.

The companion Space, TempEdge Explorer, lets you browse the test set, take a quiz, and score a predictions file entirely in the browser, with no model and no backend.

Limitations

• Business-day families cover 19 countries whose calendars are maintained in the holidays package. China is excluded because make-up workdays (调休) are not represented; the UAE is excluded because its weekend moved from Friday–Saturday to Saturday–Sunday in 2022 while the package reports a single weekend per country for all years.
• Future timezone transitions follow the current tzdata projections and can change if a government amends its rules.
• Leap seconds are out of scope; the dataset uses POSIX-style UTC arithmetic.
• Dates outside 1900–2100 are not included for the proleptic Gregorian weekday family, and timezone tasks outside 1970–2035 avoid pre-standardization oddities such as second-level local mean time.
• Prompts are template-generated. They test computation, not natural-language robustness; paraphrase testing is a separate concern.

Publishing

pip install --break-system-packages huggingface_hub
export HF_TOKEN=your_write_token
python scripts/push_to_hub.py \
  --repo-id NoahIsARider/temporaledge \
  --space-id NoahIsARider/temporaledge-explorer

The script creates both repositories if needed, uploads the dataset folder, and uploads the static Space from ../temporaledge-space.

License

Data and prompts: CC BY 4.0. IANA tzdata is public domain. The holidays package is MIT licensed.

Citation

@dataset{temporaledge2026,
  title  = {TempEdge: verified local-time and calendar arithmetic},
  author = {NoahIsARider},
  year   = {2026},
  note   = {10,900 machine-verified bilingual temporal reasoning questions},
  url    = {https://huggingface.co/datasets/NoahIsARider/temporaledge}
}