task_id stringlengths 12 20 | family stringlengths 9 17 | difficulty int64 1 4 | domain stringclasses 8
values | summary stringlengths 56 111 | tools listlengths 4 6 | in_submission bool 2
classes | split stringclasses 2
values | calibration dict | example_seed int64 0 0 | example_prompt stringlengths 564 1.09k | example_oracle stringlengths 54 1.89k |
|---|---|---|---|---|---|---|---|---|---|---|---|
loan_schedule-d1 | loan_schedule | 1 | finance | Amortization with extra payments, rate resets, deferrals and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Follow rules.md for this loan. Report the state after month 8.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}}... | [{"tool": "submit", "args": {"answer": {"balance": 149934.99, "interest": 8107.47}}}] |
loan_schedule-d2 | loan_schedule | 2 | finance | Amortization with extra payments, rate resets, deferrals and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | Follow rules.md for this loan with the applied extra payments and the rate change. Report the state after month 37.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "pyth... | [{"tool": "submit", "args": {"answer": {"balance": 174930.89, "interest": 24557.91}}}] |
loan_schedule-d3 | loan_schedule | 3 | finance | Amortization with extra payments, rate resets, deferrals and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | Simulate this loan to payoff following rules.md: applied extra payments, the rate change, deferred months and the memo's corrections.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line st... | [{"tool": "submit", "args": {"answer": {"payoff_month": 347, "total_interest": 118762.87}}}] |
loan_schedule-d4 | loan_schedule | 4 | finance | Amortization with extra payments, rate resets, deferrals and corrections | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Simulate this loan to payoff following rules.md: applied extra payments, the rate change, deferred months and the memo's corrections.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line st... | [{"tool": "submit", "args": {"answer": {"payoff_month": 171, "total_interest": 153439.87}}}] |
fx_ledger-d1 | fx_ledger | 1 | finance | Multi-currency ledger with reversals, missing FX dates, messy formats and corrections | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Apply rules.md to the ledger. Account OPS-01, as of 2026-03-29.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}... | [{"tool": "submit", "args": {"answer": {"balance_usd": 9562.38}}}] |
fx_ledger-d2 | fx_ledger | 2 | finance | Multi-currency ledger with reversals, missing FX dates, messy formats and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Apply rules.md to the ledger. Account OPS-01, as of 2026-04-10.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}... | [{"tool": "submit", "args": {"answer": {"balance_usd": 14903.77}}}] |
fx_ledger-d3 | fx_ledger | 3 | finance | Multi-currency ledger with reversals, missing FX dates, messy formats and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Apply rules.md to the ledger. Account OPS-02, as of 2026-04-02.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}... | [{"tool": "submit", "args": {"answer": {"balance_usd": 8143.61, "top_currency": "JPY"}}}] |
fx_ledger-d4 | fx_ledger | 4 | finance | Multi-currency ledger with reversals, missing FX dates, messy formats and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Apply rules.md to the ledger. Account OPS-02, as of 2026-04-13.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}... | [{"tool": "submit", "args": {"answer": {"balance_usd": 157381.01, "top_currency": "JPY"}}}] |
xnpv_xirr-d1 | xnpv_xirr | 1 | finance | Irregular-date XNPV and XIRR with mixed units, project ranking and corrections | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Value the project in cashflows.csv following rules.md, discount rate 11.5%.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standar... | [{"tool": "submit", "args": {"answer": {"npv": -398013.55}}}] |
xnpv_xirr-d2 | xnpv_xirr | 2 | finance | Irregular-date XNPV and XIRR with mixed units, project ranking and corrections | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Value the project in cashflows.csv following rules.md, discount rate 7.0%.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard... | [{"tool": "submit", "args": {"answer": {"npv": 216199.38, "irr_percent": 15.8814}}}] |
xnpv_xirr-d3 | xnpv_xirr | 3 | finance | Irregular-date XNPV and XIRR with mixed units, project ranking and corrections | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Use only the base scenario in cashflows.csv. Discount rate 6.25%. Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 ... | [{"tool": "submit", "args": {"answer": {"best_project": "CEDAR", "best_npv": 33146.29, "best_irr_percent": 7.4571}}}] |
xnpv_xirr-d4 | xnpv_xirr | 4 | finance | Irregular-date XNPV and XIRR with mixed units, project ranking and corrections | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Use only the base scenario in cashflows.csv. Discount rate 9.25%. Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 ... | [{"tool": "submit", "args": {"answer": {"best_project": "ALDER", "best_npv": 32847.77, "best_irr_percent": 12.6256}}}] |
reaction_kinetics-d1 | reaction_kinetics | 1 | natural_science | Order and rate-constant fits, Arrhenius activation energy | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Determine the kinetics of the reaction in measurements.csv following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 stan... | [{"tool": "submit", "args": {"answer": {"order": 1, "k": 0.002945844787683232}}}] |
reaction_kinetics-d2 | reaction_kinetics | 2 | natural_science | Order and rate-constant fits, Arrhenius activation energy | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 3,
"selected": false
} | 0 | Determine the kinetics of the reaction in measurements.csv following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 stan... | [{"tool": "submit", "args": {"answer": {"order": 1, "k": 0.0007432756998708229, "t_target_s": 2877.33635021043}}}] |
reaction_kinetics-d3 | reaction_kinetics | 3 | natural_science | Order and rate-constant fits, Arrhenius activation energy | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | The same reaction was run at two temperatures (run_A.csv, run_B.csv, conditions.txt). Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": ... | [{"tool": "submit", "args": {"answer": {"order": 0, "ea_kj_per_mol": 62.40909386878543}}}] |
reaction_kinetics-d4 | reaction_kinetics | 4 | natural_science | Order and rate-constant fits, Arrhenius activation energy | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | The same reaction was run at two temperatures (run_A.csv, run_B.csv, conditions.txt). Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": ... | [{"tool": "submit", "args": {"answer": {"order": 1, "ea_kj_per_mol": 48.16064437489526}}}] |
projectile_ode-d1 | projectile_ode | 1 | natural_science | Euler integration of a projectile with quadratic drag and wind | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | Simulate the projectile in launch.txt following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} run... | [{"tool": "submit", "args": {"answer": {"range_m": 134.10369780288735}}}] |
projectile_ode-d2 | projectile_ode | 2 | natural_science | Euler integration of a projectile with quadratic drag and wind | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Simulate the projectile in launch.txt following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} run... | [{"tool": "submit", "args": {"answer": {"range_m": 224.91802888903007, "flight_time_s": 8.424736796213889, "max_height_m": 92.91330493394872}}}] |
projectile_ode-d3 | projectile_ode | 3 | natural_science | Euler integration of a projectile with quadratic drag and wind | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Simulate the projectile in launch.txt following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} run... | [{"tool": "submit", "args": {"answer": {"range_m": 60.62451764678643, "flight_time_s": 3.4252231060198692}}}] |
projectile_ode-d4 | projectile_ode | 4 | natural_science | Euler integration of a projectile with quadratic drag and wind | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Simulate the projectile in launch.txt following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} run... | [{"tool": "submit", "args": {"answer": {"range_m": 153.97174446326173, "flight_time_s": 8.01209281747595}}}] |
stoichiometry-d1 | stoichiometry | 1 | natural_science | Limiting reagents, yields, purity and two-step synthesis | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Work out the batch in batch.csv for the reaction in reaction.txt, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python... | [{"tool": "submit", "args": {"answer": {"limiting": "O2", "theoretical_g": 27.61426964185262}}}] |
stoichiometry-d2 | stoichiometry | 2 | natural_science | Limiting reagents, yields, purity and two-step synthesis | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Work out the batch in batch.csv for the reaction in reaction.txt, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python... | [{"tool": "submit", "args": {"answer": {"limiting": "O2", "theoretical_g": 91.5442183326458, "percent_yield": 67.06048849193961}}}] |
stoichiometry-d3 | stoichiometry | 3 | natural_science | Limiting reagents, yields, purity and two-step synthesis | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | Two-step synthesis (reaction.txt, batch.csv, process.txt). Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standar... | [{"tool": "submit", "args": {"answer": {"limiting_step2": "NaCl", "product_g": 132.00072479512832}}}] |
stoichiometry-d4 | stoichiometry | 4 | natural_science | Limiting reagents, yields, purity and two-step synthesis | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | Two-step synthesis (reaction.txt, batch.csv, process.txt). Follow rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standar... | [{"tool": "submit", "args": {"answer": {"limiting_step2": "H2", "product_g": 15.695679212363796}}}] |
constrained_count-d1 | constrained_count | 1 | math | Counting integers under modular, digit-sum and forbidden-digit constraints | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Count the integers described in spec.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the folder ... | [{"tool": "submit", "args": {"answer": {"count": 1503}}}] |
constrained_count-d2 | constrained_count | 2 | math | Counting integers under modular, digit-sum and forbidden-digit constraints | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Count the integers described in spec.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the folder ... | [{"tool": "submit", "args": {"answer": {"count": 81283}}}] |
constrained_count-d3 | constrained_count | 3 | math | Counting integers under modular, digit-sum and forbidden-digit constraints | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 1,
"selected": true
} | 0 | Count the integers described in spec.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the folder ... | [{"tool": "submit", "args": {"answer": {"count": 20734633862}}}] |
constrained_count-d4 | constrained_count | 4 | math | Counting integers under modular, digit-sum and forbidden-digit constraints | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 1,
"selected": true
} | 0 | Count the integers described in spec.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the folder ... | [{"tool": "submit", "args": {"answer": {"count": 18923708723}}}] |
graph_paths-d1 | graph_paths | 1 | math | Shortest routes with mixed time formats, two-way roads, waypoints and route counting | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 2,
"selected": true
} | 0 | edges.csv lists one-way roads. Routes go from N10 to N4.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs... | [{"tool": "submit", "args": {"answer": {"cost": 19}}}] |
graph_paths-d2 | graph_paths | 2 | math | Shortest routes with mixed time formats, two-way roads, waypoints and route counting | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | edges.csv lists roads with travel times. Two-way roads can be used in both directions; closed roads cannot be used. Routes go from N53 to N2.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from... | [{"tool": "submit", "args": {"answer": {"cost": 142, "cost_via": 244}}}] |
graph_paths-d3 | graph_paths | 3 | math | Shortest routes with mixed time formats, two-way roads, waypoints and route counting | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 1,
"selected": true
} | 0 | edges.csv lists roads with travel times. Two-way roads can be used in both directions; closed roads cannot be used. Routes go from N79 to N97.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines fro... | [{"tool": "submit", "args": {"answer": {"cost": 330, "num_cheapest_routes": 3}}}] |
graph_paths-d4 | graph_paths | 4 | math | Shortest routes with mixed time formats, two-way roads, waypoints and route counting | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | edges.csv lists roads with travel times. Two-way roads can be used in both directions; closed roads cannot be used. Routes go from N78 to N55.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines fro... | [{"tool": "submit", "args": {"answer": {"cost": 195, "num_cheapest_routes": 4}}}] |
recurrence_mod-d1 | recurrence_mod | 1 | math | Linear and coupled recurrences mod M, 1-indexed variants, astronomically large N | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Evaluate the sequence defined in sequence.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the fo... | [{"tool": "submit", "args": {"answer": {"a_N": 258477994}}}] |
recurrence_mod-d2 | recurrence_mod | 2 | math | Linear and coupled recurrences mod M, 1-indexed variants, astronomically large N | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Evaluate the sequence defined in sequence.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the fo... | [{"tool": "submit", "args": {"answer": {"a_N": 533672609, "S_N": 590403267}}}] |
recurrence_mod-d3 | recurrence_mod | 3 | math | Linear and coupled recurrences mod M, 1-indexed variants, astronomically large N | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Evaluate the sequences defined in sequence.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the f... | [{"tool": "submit", "args": {"answer": {"a_N": 404025641, "b_N": 389477736}}}] |
recurrence_mod-d4 | recurrence_mod | 4 | math | Linear and coupled recurrences mod M, 1-indexed variants, astronomically large N | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 1,
"selected": true
} | 0 | Evaluate the sequences defined in sequence.txt.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} runs in the f... | [{"tool": "submit", "args": {"answer": {"a_N": 975929742, "b_N": 423010451}}}] |
code_repair-d1 | code_repair | 1 | software_engineering | Fix seeded bugs in a small Python module against a spec and tests | [
"list",
"read",
"edit",
"write",
"run_tests",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | lib.py has 1 bug: some functions do not match SPEC.md. Fix them without breaking the others. Hidden tests decide the grade.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "write", "args": {"path": "lib.py", "content": "import datetime as dt\n\n\ndef business_days(start, end, holidays=()):\n day = dt.date.fromisoformat(start)\n stop = dt.date.fromisoformat(end)\n off = {dt.date.fromisoformat(h) for h in holidays}\n count = 0\n while day < stop:\n if day.we... |
code_repair-d2 | code_repair | 2 | software_engineering | Fix seeded bugs in a small Python module against a spec and tests | [
"list",
"read",
"edit",
"write",
"run_tests",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | lib.py has 1 bug: some functions do not match SPEC.md. Fix them without breaking the others. Hidden tests decide the grade.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "write", "args": {"path": "lib.py", "content": "import datetime as dt\n\n\ndef business_days(start, end, holidays=()):\n day = dt.date.fromisoformat(start)\n stop = dt.date.fromisoformat(end)\n off = {dt.date.fromisoformat(h) for h in holidays}\n count = 0\n while day < stop:\n if day.we... |
code_repair-d3 | code_repair | 3 | software_engineering | Fix seeded bugs in a small Python module against a spec and tests | [
"list",
"read",
"edit",
"write",
"run_tests",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | lib.py has 2 bugs: some functions do not match SPEC.md. Fix them without breaking the others. Hidden tests decide the grade.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"too... | [{"tool": "write", "args": {"path": "lib.py", "content": "def parse_duration(text):\n units = {\"h\": 3600, \"m\": 60, \"s\": 1}\n total, number = 0, \"\"\n for ch in text.strip().lower():\n if ch.isdigit():\n number += ch\n elif ch in units and number:\n total += int(number... |
code_repair-d4 | code_repair | 4 | software_engineering | Fix seeded bugs in a small Python module against a spec and tests | [
"list",
"read",
"edit",
"write",
"run_tests",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | lib.py has 3 bugs: some functions do not match SPEC.md. Fix them without breaking the others. Hidden tests decide the grade.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"too... | [{"tool": "write", "args": {"path": "lib.py", "content": "import datetime as dt\n\n\ndef parse_duration(text):\n units = {\"h\": 3600, \"m\": 60, \"s\": 1}\n total, number = 0, \"\"\n for ch in text.strip().lower():\n if ch.isdigit():\n number += ch\n elif ch in units and number:\n ... |
machine_schedule-d1 | machine_schedule | 1 | industrial_physical_systems | Single-machine sequencing with setups, maintenance and cancellations | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Choose the job order that minimizes total weighted tardiness, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 s... | [{"tool": "submit", "args": {"answer": {"order": ["J3", "J2", "J1", "J4"]}}}] |
machine_schedule-d2 | machine_schedule | 2 | industrial_physical_systems | Single-machine sequencing with setups, maintenance and cancellations | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Choose the job order that minimizes total weighted tardiness, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 s... | [{"tool": "submit", "args": {"answer": {"order": ["J2", "J1", "J4", "J5", "J3"]}}}] |
machine_schedule-d3 | machine_schedule | 3 | industrial_physical_systems | Single-machine sequencing with setups, maintenance and cancellations | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Choose the job order that minimizes total weighted tardiness, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 s... | [{"tool": "submit", "args": {"answer": {"order": ["J7", "J6", "J4", "J1", "J3", "J5"]}}}] |
machine_schedule-d4 | machine_schedule | 4 | industrial_physical_systems | Single-machine sequencing with setups, maintenance and cancellations | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 2,
"selected": true
} | 0 | Choose the job order that minimizes total weighted tardiness, following rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 s... | [{"tool": "submit", "args": {"answer": {"order": ["J5", "J4", "J6", "J7", "J3", "J1", "J8"]}}}] |
inbox_scheduling-d1 | inbox_scheduling | 1 | office_white_collar | Earliest meeting slot from constraints spread across emails, offsets and corrections | [
"list",
"read",
"python",
"list_emails",
"read_email",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | Schedule the meeting described in brief.txt using the constraints in the inbox. Answer in the organizer's time (UTC+01:00).
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "submit", "args": {"answer": {"start": "2026-11-02 11:00"}}}] |
inbox_scheduling-d2 | inbox_scheduling | 2 | office_white_collar | Earliest meeting slot from constraints spread across emails, offsets and corrections | [
"list",
"read",
"python",
"list_emails",
"read_email",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | Schedule the meeting described in brief.txt using the constraints in the inbox. Answer in the organizer's time (UTC+01:00).
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "submit", "args": {"answer": {"start": "2026-10-13 12:30"}}}] |
inbox_scheduling-d3 | inbox_scheduling | 3 | office_white_collar | Earliest meeting slot from constraints spread across emails, offsets and corrections | [
"list",
"read",
"python",
"list_emails",
"read_email",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 3,
"selected": false
} | 0 | Schedule the meeting described in brief.txt using the constraints in the inbox. Answer in the organizer's time (UTC+01:00).
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "submit", "args": {"answer": {"start": "2026-10-21 12:00"}}}] |
inbox_scheduling-d4 | inbox_scheduling | 4 | office_white_collar | Earliest meeting slot from constraints spread across emails, offsets and corrections | [
"list",
"read",
"python",
"list_emails",
"read_email",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 1,
"selected": false
} | 0 | Schedule the meeting described in brief.txt using the constraints in the inbox. Answer in the organizer's time (UTC+01:00).
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool... | [{"tool": "submit", "args": {"answer": {"start": "2026-10-30 10:00"}}}] |
auth_forensics-d1 | auth_forensics | 1 | cybersecurity | Find the brute-force login that succeeded with grep over sshd logs, decoys and rotation | [
"list",
"read",
"grep",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | app-2 was brute-forced over SSH and an attacker got in. The attacking source is the address(es) that failed logins for invalid user names. Find the first successful login from the attacking source.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "ar... | [{"tool": "submit", "args": {"answer": {"attacker_ip": "113.66.229.154", "user": "judy", "time": "Oct 07 14:40:31"}}}] |
auth_forensics-d2 | auth_forensics | 2 | cybersecurity | Find the brute-force login that succeeded with grep over sshd logs, decoys and rotation | [
"list",
"read",
"grep",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | web01 was brute-forced over SSH and an attacker got in. The attacking source is the address(es) that failed logins for invalid user names. Find the first successful login from the attacking source.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "ar... | [{"tool": "submit", "args": {"answer": {"attacker_ip": "114.13.193.34", "user": "ivan", "time": "Oct 07 16:09:55"}}}] |
auth_forensics-d3 | auth_forensics | 3 | cybersecurity | Find the brute-force login that succeeded with grep over sshd logs, decoys and rotation | [
"list",
"read",
"grep",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 1,
"solved": 3,
"format_failures": 0,
"selected": false
} | 0 | bastion was brute-forced over SSH and an attacker got in. The attacking source is the address(es) that failed logins for invalid user names. Find the first successful login from the attacking source. The attacker rotated addresses within one /24 network: every address in that network counts as the attacking source.
Rep... | [{"tool": "submit", "args": {"answer": {"attacker_ip": "176.147.211.84", "user": "alice", "time": "Oct 15 16:40:03"}}}] |
auth_forensics-d4 | auth_forensics | 4 | cybersecurity | Find the brute-force login that succeeded with grep over sshd logs, decoys and rotation | [
"list",
"read",
"grep",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | web01 was brute-forced over SSH and an attacker got in. The attacking source is the address(es) that failed logins for invalid user names. Find the first successful login from the attacking source. The attacker rotated addresses within one /24 network: every address in that network counts as the attacking source. Logs ... | [{"tool": "submit", "args": {"answer": {"attacker_ip": "148.30.153.124", "user": "judy", "time": "Oct 08 18:12:27"}}}] |
subtitle_sync-d1 | subtitle_sync | 1 | media_content_production | Retime subtitles from sync points, frame-rate conversion and scene inserts | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | Retime movie.srt for the release using notes.md and rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} ... | [{"tool": "submit", "args": {"answer": {"start_1": "00:02:36,226", "end_2": "00:01:16,482", "start_3": "00:02:20,578"}}}] |
subtitle_sync-d2 | subtitle_sync | 2 | media_content_production | Retime subtitles from sync points, frame-rate conversion and scene inserts | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.667,
"solved": 2,
"format_failures": 0,
"selected": true
} | 0 | Retime movie.srt for the release using notes.md and rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} ... | [{"tool": "submit", "args": {"answer": {"start_1": "00:01:50,482", "end_2": "00:01:36,764", "start_3": "00:00:30,005"}}}] |
subtitle_sync-d3 | subtitle_sync | 3 | media_content_production | Retime subtitles from sync points, frame-rate conversion and scene inserts | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | Retime movie.srt for the release using notes.md and rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} ... | [{"tool": "submit", "args": {"answer": {"start_1": "00:08:59,455", "end_2": "00:02:58,267", "start_3": "00:01:15,101"}}}] |
subtitle_sync-d4 | subtitle_sync | 4 | media_content_production | Retime subtitles from sync points, frame-rate conversion and scene inserts | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | Retime movie.srt for the release using notes.md and rules.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard library>"}} ... | [{"tool": "submit", "args": {"answer": {"start_1": "00:07:38,020", "end_2": "00:03:06,031", "start_3": "00:13:53,576"}}}] |
expense_audit-d1 | expense_audit | 1 | office_white_collar | Audit expense claims against a written policy with SQL: caps, receipts, weekends, duplicates, FX and amendments | [
"list",
"read",
"sql",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 0,
"selected": true
} | 0 | Audit the expense claims of D. Tanaka following policy.md. Tables claims, tiers and rates can be queried with sql.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "sql",... | [{"tool": "submit", "args": {"answer": {"total_usd": 1006.13, "rejected": []}}}] |
expense_audit-d2 | expense_audit | 2 | office_white_collar | Audit expense claims against a written policy with SQL: caps, receipts, weekends, duplicates, FX and amendments | [
"list",
"read",
"sql",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Audit the expense claims of B. Okafor following policy.md. Tables claims, tiers and rates can be queried with sql.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "sql",... | [{"tool": "submit", "args": {"answer": {"total_usd": 1428.79, "rejected": ["C1015", "C1016"]}}}] |
expense_audit-d3 | expense_audit | 3 | office_white_collar | Audit expense claims against a written policy with SQL: caps, receipts, weekends, duplicates, FX and amendments | [
"list",
"read",
"sql",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Audit the expense claims of E. Haddad following policy.md. Tables claims, tiers and rates can be queried with sql.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "sql",... | [{"tool": "submit", "args": {"answer": {"total_usd": 1302.28, "rejected": ["C1008", "C1011", "C1013", "C1017", "C1021", "C1022", "C1025", "C1026"]}}}] |
expense_audit-d4 | expense_audit | 4 | office_white_collar | Audit expense claims against a written policy with SQL: caps, receipts, weekends, duplicates, FX and amendments | [
"list",
"read",
"sql",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Audit the expense claims of C. Lindqvist following policy.md. Tables claims, tiers and rates can be queried with sql.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "sq... | [{"tool": "submit", "args": {"answer": {"total_usd": 1048.44, "rejected": ["C1007", "C1008", "C1015", "C1016", "C1023", "C1032", "C1035", "C1036"]}}}] |
inventory_reorder-d1 | inventory_reorder | 1 | industrial_physical_systems | Reorder decisions from stock, forecasts, lead times, open POs, pack sizes, MOQs and memos | [
"list",
"read",
"python",
"submit"
] | true | train | {
"episodes": 3,
"mean_reward": 0.333,
"solved": 1,
"format_failures": 1,
"selected": true
} | 0 | Decide which SKUs to order today and how many units, following policy.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard ... | [{"tool": "submit", "args": {"answer": {"orders": {"SKU-700": 177, "SKU-539": 363, "SKU-357": 441}}}}] |
inventory_reorder-d2 | inventory_reorder | 2 | industrial_physical_systems | Reorder decisions from stock, forecasts, lead times, open POs, pack sizes, MOQs and memos | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 1,
"selected": false
} | 0 | Decide which SKUs to order today and how many units, following policy.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard ... | [{"tool": "submit", "args": {"answer": {"orders": {"SKU-656": 280, "SKU-920": 88, "SKU-252": 361, "SKU-484": 593}}}}] |
inventory_reorder-d3 | inventory_reorder | 3 | industrial_physical_systems | Reorder decisions from stock, forecasts, lead times, open POs, pack sizes, MOQs and memos | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Decide which SKUs to order today and how many units, following policy.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard ... | [{"tool": "submit", "args": {"answer": {"orders": {"SKU-191": 134, "SKU-349": 610, "SKU-828": 732, "SKU-615": 384, "SKU-724": 616}}}}] |
inventory_reorder-d4 | inventory_reorder | 4 | industrial_physical_systems | Reorder decisions from stock, forecasts, lead times, open POs, pack sizes, MOQs and memos | [
"list",
"read",
"python",
"submit"
] | false | unused | {
"episodes": 3,
"mean_reward": 0,
"solved": 0,
"format_failures": 0,
"selected": false
} | 0 | Decide which SKUs to order today and how many units, following policy.md.
Reply with exactly one JSON object per turn. Tools:
{"tool": "list", "args": {}} list the task files
{"tool": "read", "args": {"path": "<file>", "start": 1}} show 30 lines from line start
{"tool": "python", "args": {"code": "<Python 3 standard ... | [{"tool": "submit", "args": {"answer": {"orders": {"SKU-804": 384, "SKU-558": 600, "SKU-243": 100, "SKU-502": 572, "SKU-109": 348, "SKU-720": 117}}}}] |
Verified Compute Lab (Workbench v3)
An OpenEnv environment for OpenEnv Arena: 16 procedural task families across the eight Arena domains. Each task brings its own tools, and every answer is checked by an exact verifier.
Each reset builds a fresh task: a question, a few files, tables or an inbox, and a tool list in the first observation. The action is always one JSON object:
{"tool": "read", "args": {"path": "policy.md"}}
{"tool": "sql", "args": {"query": "SELECT category, SUM(amount) FROM claims GROUP BY category"}}
{"tool": "edit", "args": {"path": "lib.py", "old": "d -= 10", "new": "d -= 9"}}
{"tool": "submit", "args": {"answer": {"orders": {"SKU-412": 240}}}}
The agent gets 12 tool calls. Code and queries run in a sandbox as an
unprivileged user, with CPU, memory and output limits; read and grep return
whole lines and say where to continue.
| Family | Domain | Tools | What it takes |
|---|---|---|---|
code_repair |
software engineering | list, read, edit, write, run_tests | fix seeded bugs; at higher levels some bugs have no failing visible test and must be found from SPEC.md |
machine_schedule |
industrial | list, read, python | optimal job order for weighted tardiness, with setups, a maintenance window and cancellations |
inventory_reorder |
industrial | list, read, python | today's purchase orders from a reorder-point policy, open POs, pack sizes, MOQs and memo changes |
inbox_scheduling |
office | list, read, python, list_emails, read_email | earliest meeting slot from constraints spread across emails, UTC offsets and corrections |
expense_audit |
office | list, read, sql (no python) | audit claims against a policy: caps by city tier, receipts, weekends, duplicates, FX, dated amendments |
auth_forensics |
cybersecurity | list, read, grep (no python) | first successful login from a brute-force source, with decoys, password-using staff, a rotating /24 and log rotation |
subtitle_sync |
media | list, read, python | retime an SRT file from sync points, frame-rate conversion and a scene insert |
loan_schedule, fx_ledger, xnpv_xirr |
finance | list, read, python | amortization with corrections, multi-currency ledgers, XNPV and XIRR |
reaction_kinetics, projectile_ode, stoichiometry |
natural science | list, read, python | fits and Arrhenius, Euler integration with drag and wind, yields |
constrained_count, graph_paths, recurrence_mod |
math | list, read, python | digit DP, shortest-path counting, matrix powers for huge N |
Task ids: <family> draws a difficulty from 1-4; <family>-d<k> pins difficulty
k. A reset without a seed derives the instance from the task id and a 10-minute
window, so the rollouts of one GRPO group (reset together) share an instance and
their rewards compare the policy rather than the luck of the draw. An explicit
seed always wins.
Reward
All-or-nothing: an answer earns 1 only when every part is right. Partial credit
let a small policy learn to guess categorical fields instead of working, so it
was removed. code_repair keeps a share-of-bugs-fixed reward, since every point
needs a real edit; breaking a function that was correct scores 0. Running out of
tool calls, or {"tool": "submit", "args": {"answer": {}}}, scores 0.
Verification
- Oracles score 1.0 for every family, difficulty and seed checked; agent-view solvers (tools and prompt only: sandboxed Python, SQL, grep output, parsed emails) solve the new families on 40+ seeds each.
- Quality gates over 100 seeds per task: the answer never appears in the prompt, and the most common answer solves at most 10% of instances. The gates removed three constant answers (the first inbox slot, the identity job order, count 0).
- 598 tests, including regressions for a sandbox launch that could deadlock
under threads (no
preexec_fn) and for reads that used to clip the middle. - The linux/amd64 image passes
openenv validate --url, and a WebSocket replay of every submitted task passes against an anonymous pull on a native x86 host.
Calibration
Qwen3.8-27B, thinking off, temperature 1.0, through Hugging Face Inference Providers, with an Arena-style loop (one JSON object per reply; anything else ends the episode) and the submitted budgets (12 tool calls, 12,288 completion and 20,480 context tokens). Three episodes per pinned task; the submission trains on the 30 tasks solved in 1 or 2 of 3, covering all 16 families and eight domains.
The reply-format line was A/B tested on 70 episodes per variant: the plain "Reply
with exactly one JSON object per turn." had 12.9% format failures, while longer
escaping advice did worse (17.1% and 24.3%). Details in format_ab.json.
Local experiments and Arena results (honest nulls)
On one RTX 4090, Qwen3-4B trained with GRPO (LoRA, a custom Unsloth trainer with shared-weight vLLM, 80 steps) learned its training families but showed no gain on held-out families beyond noise.
| Arena run | Updates | Private tasks |
|---|---|---|
| v1 (17 finance/science/math tasks) | 140 | 4/40: industrial 2, office 2 (same as untrained) |
| v2 (29 tasks, 14 families) | 110 | 3/40: software 1, industrial 2, office 0 |
Across all public runs, the best single scores come from runs that changed the model least, and runs with steady gradient signal tend to lose solves the base model had. v3 shortens episodes and tightens grading; expect a score near the untrained reference.
Links
- Image:
ghcr.io/rycerzes/openenv-arena-compute-lab:v3(linux/amd64) - Source: https://github.com/rycerzes/openenv-arena-compute-lab
- Dataset: https://huggingface.co/datasets/rycerzes/openenv-arena-compute-lab
- OpenEnv at
86a180ede21e044f7929b9a7783ad83aa67d83a3, as pinned by the Arena
Files
tasks.jsonl: one row per pinned task id (64), with its domain, tools, whether it is in the submission, its calibration, and a seed-0 example prompt and oracle.source/: environment source, tests, solvers, trainers and scripts.calibration_summary.json: per-task results against Qwen3.8-27B.format_ab.json: the reply-format A/B.
- Downloads last month
- 102