jessicayjung75/finance-ops-spreadsheet-env

Finance-ops spreadsheet agent training environment

An RL / agent training environment for spreadsheet work that a business user would hand an analyst. The agent gets a workbook (input.xlsx) and a plain-English request, and must add the requested result sheets to the same workbook. This is sold as a training environment, not as a frontier evaluation.

What's in the bundle (248 tasks)

folder	tasks	what
tasks-core/	40	8 families x 5 seeds, tidy tables: AR aging, bank reconciliation, sales commission with tiers and a new-rep ramp, budget vs actual by department, customer de-duplication, multi-currency expense conversion, inventory reorder quantities, payroll overtime
tasks-core-raw/	40	the same tasks written the way an ERP report export looks (see below)
tasks-hard/, tasks-hard-raw/	10 + 10	revenue-recognition schedules (amendments, cancellations, daily proration, monthly cent rounding with a lifetime plug) and cash application (memo invoice references, customer restrictions, oldest-invoice fallback, unapplied credit)
tasks-close/	8	a full month-end AR close over ten raw input sheets across three entities and currencies: customer merges, void invoices, bounced payments, credit memos, free-text remittance memos, cut-off, collections notes, FX and GL. The policy is in a Close Memo sheet, not the instruction. Five output sheets must all be right.
tasks-finance/, tasks-finance-raw/	20 + 20	the finance tier, controller-level work with the policy in a memo sheet and every output sheet graded all-or-nothing: a Q3 three-entity group consolidation (USD parent, GBP and a partly owned EUR subsidiary; per-entity account maps, closing vs average vs historical rates, CTA, intercompany eliminations with FX differences, investment elimination with goodwill and NCI) and a Q3 lease close (finance and operating leases, advance and arrears, initial direct costs, incentives, modifications with remeasurement, current portion), with an AP register as a distractor. Half are held out (-ho- in the name): they ship with the workbook, instruction and grader but without answers or reproducible seeds, and we grade those runs
tasks-group/, tasks-group-raw/, tasks-groupx/, tasks-groupx-raw/	10 + 10 + 10 + 10	the group close tiers: the same three-entity group closed from 700+ raw GL journal lines (posting-date cutoff, drafts, a duplicate re-import batch), with an item-by-item intercompany investigation (group_close: one pair, one item of each kind; group_close_x: both pairs, a principles-only memo and an unknown number of unrecorded, disputed, misposted, duplicate and in-transit items, credit notes and part payments), monthly average-rate translation and consolidation. Six chained output sheets graded all-or-nothing. Half are held out (-ho-) the same way as the finance tier
tasks-precedent/, tasks-precedent-raw/, tasks-precedentx/, tasks-precedentx-raw/	10 + 10 + 10 + 10	the close-by-precedent tiers: last month's finished close package is the only spec. The agent recovers the reserve rates by aging bucket, capitalization threshold, depreciation convention, revenue recognition rule, overhead allocation driver and bonus base from August, then closes September with six matching sheets (precedent: the memo gives each policy's form; precedent_x: the memo states none of them). The generator rejects any seed where August fits more than one policy. Half are held out (-ho-) the same way as the finance tier
tasks-precedentq/, tasks-precedentq-raw/	10 + 10	close by precedent over a quarter: from August's finished package alone, close September to December with four stated mid-period events (asset sale, contract cancellation, a new capitalization threshold and bonus rate from November). The four monthly summaries and December's reserve, fixed assets, deferred revenue and allocation are graded, so a wrong policy or a wrong carried balance shows in every later month. Half are held out (-ho-)
code/		the seeded generator (build.py, families.py, hard.py, close.py, finance.py, group_close.py, group_close_x.py, precedent.py, precedent_q.py, messy.py), the grader (grade.py) and the HUD / Harbor check (hud_verify.py)

Raw ERP variants have banner rows above the header, a header repeated at a page break, dates in mixed formats, money as text ($1,234.50, (1,234.50)), stray spaces in keys, ids stored as numbers in one sheet and as text in another, and subtotal and grand-total lines. The correct answer is unchanged, so every quirk is one a careful analyst handles without guessing.

New variants: python code/build.py --out tasks [--messy] [--set core|hard|close|finance|group|groupx|precedent|precedentx|precedentq] --per-family N.

Task format (Harbor) and grader

instruction.md, task.toml, environment/{Dockerfile,input.xlsx}, solution/solve.sh, and tests/{test.sh,grade.py,expected.json} (held-out tasks ship without solution/ and expected.json). The grader recalculates formulas headless with LibreOffice, so formula and value answers grade alike. Numbers must match within 0.011 and text after trimming. The reward is 1 only when every expected cell matches; the share of matching cells is written to grade.json.

Quality checks

• build.py checks every task as it builds it: the reference passes, the untouched workbook fails, and a copy with one wrong cell fails (248/248 valid in this bundle; held-out tasks were proven at build time, before their answers were removed).
• hud_verify.py checks that each task adapts cleanly through HUD's Harbor adapter, with the reference scoring 1 and doing nothing scoring 0.
• Measured difficulty: Claude Sonnet 4.5 solved core 31/31, raw 23/27 and hard 6/6; Claude Opus 5.5 solved hard 20/20. Group close tiers, on the held-out half (one run, up to 25 tool turns): group_close Opus 5.5 9/10, Sonnet 5.5 8/10, Haiku 5.5 0/10; group_close_x (on the 8 of 10 tasks a since-fixed generator bug didn't touch) Opus 8/8, Sonnet 6/8, Haiku 0/8. Close-by-precedent tiers, on the held-out half: precedent Opus 5.5 10/10 (after fixing halves-up rounding and the messy variant's banners; Sonnet 5.5 6/10 and Haiku 5.5 3/10 before those fixes); precedent_x Opus 9/10, Sonnet 5/10, Haiku 0/10. precedent_q (one run, up to 40 tool turns): Opus 5.5 6/10, Sonnet 5.5 7/10, Haiku 5.5 0/10. The month-end close and finance tiers have not been calibrated yet.

Licence and provenance

All records, names and companies are generated; there are no real people or businesses. Copyright Keeno Corp. The buyer's rights (non-exclusive, including commercial model training) are those granted in the sale agreement.