CashSage: 100,000 personal finance computations, and the ones it refuses to answer

A 100,000-row training corpus for the money questions people actually ask: how long to pay off a card, which debt to attack first, what an extra 200 dollars a month does to a mortgage, what a Social Security benefit works out to, what a 401k match is worth, and what a fund's expense ratio costs over thirty years.

What this dataset proves, and how you check it

rows	100,000, 7 families
how labels were produced	executed by a reference simulator, then checked by an independent oracle
what the simulator is validated against	the official SSA and Regulation 12 CFR 1026 appendix M2 worked examples
labels authored or judged by a language model	zero
rows that teach the model not to compute	20,000: ask for a missing input, decline, or state the question is undecidable
what it does to a model	gemma-3-4b-it: 1.3% to 100.0% exact match, held out
significance	p = 1.2e-178, exact McNemar
unseen-year test	300 rows on 2026 statutory limits the corpus never contained, scored 100.0%
decontamination	101,980 rows screened against FinQA, ConvFinQA, TatQA and FinChain, 0 flagged
what ships beside the rows	generator, oracle, full reference simulator, scorer, every eval slice, per-row verdicts

Every number above is recomputable on a laptop with stdlib Python and no GPU.

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

A fifth of the corpus is not a computation

shard	rows	what it teaches
compute	80,000	produce the answer, with the arithmetic shown
clarify	10,000	ask for the one missing input instead of assuming it
undecidable	5,000	say the question cannot be answered as posed
refuse	5,000	decline, because this is regulated advice territory

20,000 rows train the model not to answer. That is the part that decides whether a personal finance assistant is usable. A model that invents an interest rate to complete a payoff schedule, or that answers "should I roll over my 401k" as if it were arithmetic, is worse than one that says it needs more information.

Clarification is allocated where inputs genuinely go missing: card payoff and amortization take 27% each, savings timelines 23%, retirement 20%, and the two families with fixed statutory inputs take almost none.

The seven families

family	rows	closed form?
card_payoff	13,600	no
multi_debt	13,600	no
amortization_extra	13,600	no
ssa_pia	11,200	no, bend points
retirement_401k	11,200	no, vesting and match tiers
fund_cost_n1a	10,400	no, N-1A cost model
savings_timeline	6,400	yes

51% of the corpus has no closed form, meaning the answer cannot be reached by substituting into a formula and must be iterated period by period. That is deliberate, and the weighting was set by measurement rather than taste: at Day-0 the base model scored 32.1% strict on savings_timeline, the one family with a closed form, and 0.0% strict on every other family. Mass was shifted away from the family the base could already do and toward the six it could not.

Decontamination

reference set	questions
FinQA	1,147
ConvFinQA	434
TatQA	1,669
FinChain fragments	4,505

101,980 rows screened against 38,744 reference 8-grams. 0 flagged. The flagged-id list is published and it is empty, so the claim is checkable rather than asserted.

Verification

Every label is produced by executing the computation, not by writing it down. Answers are checked against an independent implementation, and a row whose two implementations disagree is dropped rather than repaired. No language model appears anywhere in the label path.

Held-out slices:

slice	rows	what it isolates
held-out	1,500-row held-out set	the competition's own yardstick
2026 rules	300-row unseen-rule-year slice	statutory limits the training corpus never saw
hard	180-probe hard shard	the worst cases across all seven families
Day-0	160 hand-checked probes	the pre-registered base-model gate

The 2026 slice is the interesting one. Contribution limits, bend points and tax brackets change every year, so it tests whether the model applies a rule it was given or recites one it memorized.

Adaptive Data

A per-row merge was pre-registered before the pass ran (data/raw/cashsage/adaption/PER_ROW_MERGE_PREREGISTRATION.md), specifying exactly which fields enhancement was allowed to touch and what would cause a row to be rejected. Enhancement applies to instruction sentences only: never a computed value, never a rate, never a statutory limit, never a date. A rewritten instruction cannot corrupt an executed answer.

Writing the acceptance rule down before seeing the returned rows is the whole point. A merge rule authored after inspecting the output is a rule fitted to the output.

Token budget, measured

Measured with the real gemma-3-4b-it tokenizer before training. The p99 is 1,309 tokens, so training uses a 1536 window. The spec had assumed 1280, which would have truncated 13.4% of rows, and under completion-only masking a truncated completion is a silently poisoned label. Measuring caught it.

Try it, and everything that backs it

Live side by side demo: https://huggingface.co/spaces/Jainamshahhh/cashsage-demo Enter your own input and watch the base model and CashSage-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/cashsage-4b	https://www.kaggle.com/datasets/jainamshahhh/cashsage-4b
dataset	https://huggingface.co/datasets/Jainamshahhh/cashsage	https://www.kaggle.com/datasets/jainamshahhh/cashsage-personal-finance-computation

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 personal finance conventions: IRS contribution limits, Social Security bend points, N-1A fund cost disclosure. Not tax preparation, not state-specific rules, not securities selection. This is not financial advice, and 5,000 rows exist specifically to teach the model to say so.