Datasets:

nice-bill
/

shiftdesk

task_id stringlengths 11 17	domain stringclasses 8 values	family stringclasses 8 values	level int64 1 3	budget_calls int64 6 12	example_question stringlengths 75 361	answer_format stringlengths 8 89	files listlengths 0 3
module-repair-L1	software-engineering	module-repair	1	8	utils.py has 1 bug (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. Submit the complete corrected contents of utils.py as the answer string.	a string holding the whole fixed file, with newlines escaped as \n inside the JSON string	[ "test_utils.py", "utils.py" ]
module-repair-L2	software-engineering	module-repair	2	10	utils.py has 2 bugs (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. Submit the complete corrected contents of utils.py as the answer string.	a string holding the whole fixed file, with newlines escaped as \n inside the JSON string	[ "test_utils.py", "utils.py" ]
module-repair-L3	software-engineering	module-repair	3	12	utils.py has 3 bugs (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. The visible tests do not cover every bug. Submit the complete corrected contents of utils.py as the answer string.	a string holding the whole fixed file, with newlines escaped as \n inside the JSON string	[ "test_utils.py", "utils.py" ]
line-oee-L1	industrial-physical-systems	line-oee	1	8	From states.csv and notes.txt: what was the availability of machine M36 this shift, as a percent of the 480 planned minutes spent in RUN? Round to 1 decimal.	a number such as 73.5	[ "notes.txt", "states.csv" ]
line-oee-L2	industrial-physical-systems	line-oee	2	10	What was the OEE of machine M58 this shift, in percent? Round to 1 decimal.	a number such as 61.2	[ "notes.txt", "production.csv", "states.csv" ]
line-oee-L3	industrial-physical-systems	line-oee	3	12	Which machine had the lowest OEE this shift, and what was it? Answer as the machine id, a space, and the OEE percent rounded to 1 decimal.	a string such as "M17 54.3"	[ "notes.txt", "production.csv", "states.csv" ]
kinetics-lab-L1	natural-science	kinetics-lab	1	8	From run.csv and protocol.txt, what is the rate constant k in 1/min? Give 4 significant figures.	a number such as 0.04213	[ "protocol.txt", "run.csv" ]
kinetics-lab-L2	natural-science	kinetics-lab	2	10	From run.csv and protocol.txt, what is the half-life in minutes? Give 4 significant figures.	a number such as 17.42	[ "protocol.txt", "run.csv" ]
kinetics-lab-L3	natural-science	kinetics-lab	3	12	From run.csv and protocol.txt: what is the reaction order and the half-life in minutes? Answer as the order, a space, and the half-life to 4 significant figures.	a string such as "1 23.45"	[ "protocol.txt", "run.csv" ]
expense-policy-L1	office-white-collar	expense-policy	1	8	Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.	a number such as 812.40	[ "expenses.csv", "policy.md" ]
expense-policy-L2	office-white-collar	expense-policy	2	10	Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.	a number such as 812.40	[ "expenses.csv", "policy.md", "rates.txt" ]
expense-policy-L3	office-white-collar	expense-policy	3	12	Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.	a number such as 812.40	[ "expenses.csv", "policy.md", "rates.txt" ]
loan-desk-L1	finance-economics	loan-desk	1	8	Read quote.txt. What is the fixed monthly payment in dollars? Round to cents.	a number such as 1234.56	[ "quote.txt" ]
loan-desk-L2	finance-economics	loan-desk	2	10	Use loans.csv and rules.txt. What is the outstanding balance of loan LN-5969 right after its 51th scheduled payment? Dollars, rounded to cents.	a number such as 98765.43	[ "loans.csv", "rules.txt" ]
loan-desk-L3	finance-economics	loan-desk	3	12	Use loans.csv and rules.txt. With the extra payments, how much total interest is paid on loan LN-4578 until it is fully repaid? Dollars, rounded to cents.	a number such as 12345.67	[ "loans.csv", "rules.txt" ]
exact-count-L1	mathematics-or-formal-reasoning	exact-count	1	6	How many integers x with 1 <= x <= 1285975 are divisible by 11 or by 15, but not divisible by 3?	an integer	[]
exact-count-L2	mathematics-or-formal-reasoning	exact-count	2	8	grid.txt is a map: '.' is open, '#' is blocked. Moving only right or down one cell at a time, how many paths go from the top-left cell to the bottom-right cell without entering a blocked cell? Give the exact integer (0 if none).	an integer	[ "grid.txt" ]
exact-count-L3	mathematics-or-formal-reasoning	exact-count	3	8	Count the 49-digit positive integers (no leading zero) whose digits sum to 292 and that never use the digit 6. Give the count modulo 1000000007.	an integer	[]
auth-hunt-L1	cybersecurity	auth-hunt	1	8	Which source IP has the most failed password attempts in auth.log? Answer with the IP address. The log is auth.log.	a string	[ "auth.log" ]
auth-hunt-L2	cybersecurity	auth-hunt	2	10	Find the compromise: a source IP with at least 5 failed passwords for one user within any 10-minute span, followed by an accepted login from that same IP for that same user no more than 10 minutes after its last failure. Exactly one IP matches. Answer with that IP. The log is auth.log.	a string	[ "auth.log" ]
auth-hunt-L3	cybersecurity	auth-hunt	3	12	An attacker rotated source addresses inside one /24 network. Treat all IPs in the same /24 as one source. Find the /24 with at least 5 failed passwords for one user within any 10-minute span, followed within 10 minutes of its last failure by an accepted login for that same user from the same /24. Answer in CIDR form, l...	a string	[ "auth.log" ]
subtitle-qc-L1	media-content-production	subtitle-qc	1	8	episode.srt is a subtitle file; each line starts with the speaker's name. What is the total time, in seconds, that ANA's subtitles are on screen? Give 3 decimals.	a number such as 41.250	[ "episode.srt" ]
subtitle-qc-L2	media-content-production	subtitle-qc	2	10	Apply notes.txt to episode.srt. What is the end timestamp of the last remaining cue after the edit? Answer in SRT form HH:MM:SS,mmm.	a string such as "00:03:12,480"	[ "episode.srt", "notes.txt" ]
subtitle-qc-L3	media-content-production	subtitle-qc	3	12	Check episode.srt against style.txt. Which cue numbers break the reading-speed rule? Answer with the cue numbers in increasing order, separated by commas (empty string if none).	a string such as "3,8,15"	[ "episode.srt", "style.txt" ]

ShiftDesk

An OpenEnv environment for OpenEnv Arena. 24 tasks: 8 families (one per arena domain) x 3 levels. Every reset generates a fresh instance (files, question, answer) from a seeded generator, so there are no fixed answer keys and this dataset holds none.

Domain	Family	What the agent does
Software engineering	module-repair	Fix 1 to 3 seeded bugs in utils.py; submit the whole file as a JSON string; hidden randomized tests against the docstring spec
Industrial & physical	line-oee	Availability and OEE from a machine state-change log, production counts and breaks
Natural science	kinetics-lab	Rate constant, half-life and reaction order from noisy concentration data with QC-failed rows and unit changes
Office & white-collar	expense-policy	Reimbursable total under a written policy: caps, receipts, currencies, duplicates
Finance & economics	loan-desk	Annuity payment, balance after k payments, interest with extra payments
Math & formal reasoning	exact-count	Inclusion-exclusion counts, lattice paths on a blocked grid, digit DP modulo 1e9+7
Cybersecurity	auth-hunt	Find the brute-force that succeeded in an sshd log; decoys; /24 rotation at level 3
Media & content	subtitle-qc	SRT on-screen time, edit notes with shifts and cuts, reading-speed violations

What it teaches

Read the tool list, don't memorize it. Tool names and argument names are drawn per instance (for example exec_code with code, or sandbox with program; final_answer with value). The reset text lists them with one example call each. An unknown tool costs a call and returns the valid names.
One valid JSON object per turn, including code. The Python tool takes source as a normal JSON string, and the software task needs a whole file as the answer, so the policy has to escape newlines and quotes correctly.
Use tools for data, not eyes. Reads are capped at 2,500 characters; logs run to ~10,000, so the policy has to compute in Python.

Reward

1.0 for a correct final answer (numeric tolerance stated per family, exact match for ids, sets and timestamps), otherwise 0.0. No answer, running out of calls, or an invalid reply scores 0.0.

Same instance for a group

Rollouts that reset the same task within the same 5-minute window get the same instance (seeded from a secret in the image), so the five parallel rollouts of an optimizer step are compared on one problem.

Checks

Reference answers score 1 and plausible wrong answers score 0 on hundreds of seeds per task; independent solvers that read only the files reproduce the labels for the log, OEE and subtitle families; every bug mutation is detectable by the hidden tests; openenv validate passes; every task replays to a terminal reward over WebSocket. Policy code runs as nobody with CPU, memory and process limits.

Source: https://github.com/nice-bills/shiftdesk

Downloads last month: 15

Number of rows:

Total file size:

15 kB