piyushgupta53/formulagym-arena

FormulaGym for OpenEnv Arena

An OpenEnv environment where an agent repairs formulas in real business spreadsheets from the public Enron corpus. It practises what office and finance work needs from a spreadsheet: read a workbook with tools, find the rows and columns that matter, and write one Excel formula that is right for every row it is copied to.

This dataset describes the 50 tasks submitted to OpenEnv Arena. What runs is the Docker image named in the submission; this card and tasks.jsonl say what is in it and how it was chosen.

Task types

Each task is one formula block (a run of cells sharing one formula shape, e.g. a column of row totals) in one workbook. The same block can be posed three ways:

Type	task_id	What the agent sees	What it submits
rebuild	fg-…	the block's cells are empty	the formula for the top-left cell
repair	fg-….repair	the block holds a wrong formula (a shifted row, a dropped term, a range end off by one, a swapped sign, a wrong lookup column) and the wrong values it gives	the correct formula
pasted	fg-….pasted	one cell of the block holds its typed-in value instead of its formula; only the sheet is named	the cell and its formula

Tools and budget

One JSON object per turn: list_sheets, read_range (≤ 400 cells, shown as ADDRESS=value [=FORMULA]), find (text search), try_formula (the value a formula gives right now; it does not say whether it is right) and submit. 16 actions in all; the last must be submit. Observations are capped at 3,000 characters.

Reward (deterministic, no LLM)

LibreOffice recalculates the workbook with the submitted formula filled across the block, on the real inputs and on perturbed copies of them (random changes, one input at a time, and edge values such as 0 and large values of the other sign). The reward is 1 if the outputs match the original formula's outputs on every copy, else 0.

• A typed-in number fails, because the outputs move when the inputs move.
• Any equivalent formula passes (=A1+A2+A3 and =SUM(A1:A3) alike).
• An answer identical to the original (ignoring spaces, $ and a leading +) scores 1 without recalculation.
• Repair: submitting the given wrong formula scores 0. Pasted: the right formula in the wrong cell scores 0.

How the 50 tasks were chosen

Step	Count
Formula tasks built from 20,872 Enron .xls files (filters for errors, volatile functions, trivial formulas, personal data, near-duplicate workbooks)	15,921 tasks, 2,742 workbooks
Pool: the training splits only (never the held-out test/eval workbooks or their near-copies)	12,223
Dropped: the same formula shape visible on another sheet (1,541), workbook data over 1 MB (1,488), a strong model failed it (266), uses a named range (175)	8,753 eligible
Seeded draw (seed 7): family quotas tilted to lookups, conditional sums and finance; a mix of tiers; ≤ 1 task per workbook (≤ 2 for rare families)	330 tasks
Checks inside the image, per type: answer hidden; original → 1; typed number, given wrong formula, wrong cell, empty submit → 0; the checker rejects generated wrong formulas; grading under 60 s	725 of 769 task ids pass
Screening: Qwen/Qwen3.8-27B (revision 1d4bf0f, the arena's base model), thinking off, the arena's one-JSON-object protocol, 4–5 tries per task id (3,132 tries)	46% of tries solved; 29% ended on a reply that was not a single JSON object
Chosen: 45 tasks solved in some tries but not all, with at least one failure that was a wrong answer (not only a format failure), ranked by how mixed the results were; plus 5 the 27B never solved but a stronger model did. One task per workbook, ≤ 8 per formula family	50

The 50: 26 rebuild, 14 repair, 10 pasted · aggregate 8, arithmetic 8, cross-sheet 8, logic 8, lookup 6, conditional sums 5, finance 4, dates 3 · tier (by functions used) hard 20, medium 24, easy 6 · mean screening solve rate 0.46.

The screening run went on after the choice; tasks.jsonl reports each task's solved/tries from the fuller results (5–6 tries per task). With them, all 45 mixed tasks are still mixed (2/5 to 5/6), and the 5 hard ones have 0 or 1 solves.

Files

• tasks.jsonl: one line per task: task_id, type, family, tier, screen_solved, screen_tries, workbook (path in the SheetJS/enron_xls mirror), sheet, and for rebuild and repair the target range. Answers (the original formulas, the pasted cell) are not included.
• source/: the environment and checker code (OpenEnv server, task types, checks, screening and selection scripts).

Limits

• The checker perturbs numeric inputs only. Formulas that differ only in how they treat text or blank cells, or in rows the workbook leaves empty, can pass; tasks where the checker rejected none of the generated wrong formulas were dropped, but this is a test on generated mistakes, not a proof.
• Grading uses LibreOffice, not Excel; a few functions differ in edge cases.
• Screening used 4–5 tries per task, so solve rates are rough. Our server and prompt may differ from the arena's harness (its system prompt and sampling settings are not published).
• Personal data was filtered with keyword rules, a model check and manual review of samples; names of Enron employees in work roles were kept. Report anything you find and it will be removed.

Source and license

Spreadsheets: the Enron corpus as mirrored by SheetJS/enron_xls (CC0 1.0), made public during the FERC investigation. This dataset: CC0 1.0. The code in source/: Apache-2.0.