ChronoCalc: 60,000 date and time computations, every answer solved twice

A 60,000-row training corpus for the calendar arithmetic that real systems get wrong: counting business days across federal holidays, rolling a date off a weekend, converting across timezones through a daylight saving transition, finding the next occurrence of a recurring schedule, and computing ISO week numbers including the years that have a week 53.

What this dataset proves, and how you check it

rows	60,000, 5 families, 51 distinct templates
how labels were produced	solved twice, by two independently written solvers
solver disagreements	0 across all 60,000 rows, and a disagreeing row is dropped, never repaired
labels authored or judged by a language model	zero
reproducibility	the corpus regenerates byte for byte from one seed; a row is a pure function of its integer id
what it does to a model	gemma-3-4b-it: 11.0% to 91.0% exact match, held out
significance	p = 7.2e-113, exact McNemar
what this card volunteers	the base is already 69 to 86% at plain day-walking, and wins 13 held-out rows outright
what ships beside the rows	both solvers, the scorer, every eval slice, per-row verdicts

Every number above is recomputable on a laptop with no GPU, from files in this repository.

Built for the Adaption AutoScientist Challenge, Part 2 (Math and Code), and co-optimized with Adaptive Data (Adaption Labs).

Every gold answer is computed twice, by two solvers that share no code

solver_a produces the answer and the line-by-line working. solver_b is written from scratch with different primitives and produces the answer again. A row whose solvers disagree is dropped and logged, never repaired.

On the released corpus there are 0 dual-solver disagreements across 60,000 rows. No language model appears anywhere in the label path.

A row is also a pure function of its integer id, so regenerating from the same preset reproduces the file byte for byte. Ids come from a shared allocator that hands each split a disjoint range, so a training row and a held-out row cannot collide by construction rather than by careful shuffling.

The five families

family	rows	what it computes
F1	9,802	business days forward or backward across federal holidays
F2	8,614	rolling a date off a weekend or holiday, forward or backward
F3	14,554	elapsed time across timezones and daylight saving transitions
F4	12,178	the next occurrence of a recurring schedule
F5	14,852	ISO week numbers, including cross-year week 53

Why guessing does not work here

The strongest property of this corpus is how thin the answer distribution is. If a model could score by memorizing a common answer, the most frequent answer would carry real mass. It does not:

family	distinct answers	share held by the single most common answer
F1	2,513	0.1%
F2	2,035	0.2%
F3	851	0.3%
F4	1,983	0.7%
F5	562	0.6%

The best possible constant guess scores under one percent. Every point has to be computed.

The traps are also load-bearing rather than incidental:

• 67.9% of F1 rows are ones where a holiday actually changes the answer, and only 0.7% are the degenerate case where the answer equals the anchor plus N.
• 72.7% of F2 rows genuinely require an adjustment.
• Of 5,824 same-zone F3 rows, 5,717 span a daylight saving transition. 42 distinct IANA zones appear.
• 64.6% of F4 rows are ones where the winning occurrence moved.
• 2.6% of F5 rows cross a year boundary, across 393 distinct cross-year answers, which is where ISO week numbering actually bites.

Completions are long on purpose

Every completion enumerates its computation line by line before emitting a single <answer> tag.

This is a deliberate cost. A terser format would score identically on exact match and be worth less, and this project has twice trained a model that became terser than its base and then lost win rate on tasks it was never trained on, because terseness generalizes out of the shard that taught it. Showing the work is also what makes a wrong answer diagnosable rather than merely wrong.

What the prompt promises

Every convention needed to compute the answer is stated in the prompt: the anchor date in ISO 8601 with its weekday named, the counting rule, the rolling rule, IANA zone names wherever a timezone appears, and the full schedule rule for recurring events.

One disclosed exception. In 90% of rows the holiday calendar is named by statute rather than enumerated, because knowing the US federal calendar is part of the behaviour being trained. The other 10% print the calendar in full, and a dedicated 200-row enumerated held-out slice exists so the headline claim can be rechecked with the calendar fully disclosed. That slice is the answer to "is it computing, or does it just know the holidays?"

Held-out slices

Disjoint from training on id, prompt and scenario:

slice	rows	what it isolates
headline	500-row held-out set	the competition's own yardstick
hard	300-row hard shard, 60 per family	the hardest case in every family
enumerated	200-row enumerated slice	calendar printed in full, no statute knowledge needed
range	150-row range slice, 2031 to 2035	dates beyond the training window

The range slice matters: it tests dates outside the period the training corpus covers, so a model cannot pass it by having seen nearby years.

Verification

• 29 of 29 integrity checks pass on the corpus, and 24 of 24 on the enumerated slice, which has no F3 or F5 rows by design.
• 0 dual-solver disagreements across 60,000 rows.
• 0 decontamination overlaps.
• 21 file pairs proven disjoint on id, prompt and scenario.
• 51 distinct template ids, so phrasing is not a single mould.

Token budget, measured

Measured with the real gemma-3-4b-it tokenizer rather than assumed:

	tokens
prompt p50 / p99 / max	331 / 613 / 766
completion p50 / p99 / max	182 / 456 / 521
total p50 / p90 / p99 / max	532 / 701 / 833 / 960

Training uses max_len 1024 and 0 of 60,000 rows exceed the budget. A 768 window would truncate roughly the top decile, and a truncated completion loses its <answer> tag entirely, so the row would teach nothing under completion-only masking.

Adaptive Data

An enhancement pass was run and refused. The returned rows rewrote values the labels depend on, including dates inside the question, so merging them would have produced confidently wrong gold answers. The pass was aborted and the decision is documented in docs/adaption/chrono_enhancement_verdict.md rather than quietly dropped.

Instruction sentences remain the only surface any enhancement is permitted to touch, because a rewritten instruction cannot corrupt a computed answer while a rewritten date can.

Try it, and everything that backs it

Live side by side demo: https://huggingface.co/spaces/Jainamshahhh/chronocalc-demo Enter your own input and watch the base model and ChronoCalc-4B answer it under identical greedy decoding. The GPU backend scales to zero, so a cold first request takes about a minute.

	Hugging Face	Kaggle
model	https://huggingface.co/Jainamshahhh/chronocalc-4b	https://www.kaggle.com/datasets/jainamshahhh/chronocalc-4b
dataset	https://huggingface.co/datasets/Jainamshahhh/chronocalc	https://www.kaggle.com/datasets/jainamshahhh/chronocalc-date-time-arithmetic

Released on both platforms, with the scorer, every eval slice, the per row verdicts and significance.py alongside, so every number on this page can be recomputed rather than trusted.

Limitations

US federal holiday calendar, Gregorian dates, IANA timezone database. Not other national calendars, not historical calendar reforms, not leap seconds. Dates range across the training window with a dedicated out-of-range slice for 2031 to 2035.