v4 task set: loans finance family, protected reward
Browse files- README.md +40 -33
- data/tasks.jsonl +9 -10
README.md
CHANGED
|
@@ -7,6 +7,7 @@ tags:
|
|
| 7 |
- agents
|
| 8 |
- tool-use
|
| 9 |
- software-engineering
|
|
|
|
| 10 |
configs:
|
| 11 |
- config_name: tasks
|
| 12 |
data_files:
|
|
@@ -14,49 +15,55 @@ configs:
|
|
| 14 |
path: data/tasks.jsonl
|
| 15 |
---
|
| 16 |
|
| 17 |
-
# Workbench (
|
| 18 |
|
| 19 |
An OpenEnv environment that trains **workspace habits** in a `run_bash` / `submit` sandbox, built for [OpenEnv Arena](https://openenvarena-arena.hf.space/).
|
| 20 |
|
| 21 |
- **Source:** https://github.com/moarshy/workbench
|
| 22 |
-
- **Image:** `ghcr.io/moarshy/workbench:
|
| 23 |
-
- **Notes:** https://github.com/moarshy/openenv-arena-notes
|
| 24 |
-
|
| 25 |
-
## Tasks (
|
| 26 |
-
**
|
| 27 |
-
- **
|
| 28 |
-
-
|
| 29 |
-
-
|
| 30 |
-
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
**
|
| 34 |
-
- **
|
| 35 |
-
- **plant-identify / plant-tune.** Identify a sealed heater simulator by experiment. `tune` also requires PI tuning, and the grader re-simulates the closed loop with hidden noise.
|
| 36 |
-
- **downtime-medium.** A sensor log plus a handbook with amendments.
|
| 37 |
|
| 38 |
**Office (1)**
|
| 39 |
- **expenses-easy.** A policy audit.
|
| 40 |
|
| 41 |
-
**
|
| 42 |
-
-
|
| 43 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
## Calibration
|
| 46 |
-
Qwen3.8-27B, thinking off, 8 episodes per task, counting full solves:
|
| 47 |
-
|
| 48 |
-
| task |
|
| 49 |
-
|---|---|
|
| 50 |
-
| pyrepo-bug1 |
|
| 51 |
-
| pyrepo-
|
| 52 |
-
|
|
| 53 |
-
|
|
| 54 |
-
|
|
| 55 |
-
|
|
| 56 |
-
|
|
| 57 |
-
|
|
| 58 |
-
| downtime-medium |
|
| 59 |
-
|
|
|
|
| 60 |
|
| 61 |
## Licences
|
| 62 |
The vendored libraries keep their own licences: toolz and boltons are BSD-3-Clause, tabulate and inflection are MIT.
|
|
|
|
| 7 |
- agents
|
| 8 |
- tool-use
|
| 9 |
- software-engineering
|
| 10 |
+
- finance
|
| 11 |
configs:
|
| 12 |
- config_name: tasks
|
| 13 |
data_files:
|
|
|
|
| 15 |
path: data/tasks.jsonl
|
| 16 |
---
|
| 17 |
|
| 18 |
+
# Workbench (v4 task set: finance traps + protected reward)
|
| 19 |
|
| 20 |
An OpenEnv environment that trains **workspace habits** in a `run_bash` / `submit` sandbox, built for [OpenEnv Arena](https://openenvarena-arena.hf.space/).
|
| 21 |
|
| 22 |
- **Source:** https://github.com/moarshy/workbench
|
| 23 |
+
- **Image:** `ghcr.io/moarshy/workbench:v4`
|
| 24 |
+
- **Notes and pre-registration:** https://github.com/moarshy/openenv-arena-notes/blob/main/versions/v8-workbench-v4.md
|
| 25 |
+
|
| 26 |
+
## Tasks (9)
|
| 27 |
+
**Finance (3), new in v4**
|
| 28 |
+
- **loans-easy / medium / hard.** Month-end loan-loss provisions: stage loans by days past due, apply provision rates, write `provisions.json`.
|
| 29 |
+
- **Late caveat:** a dated risk-committee memo changes a rate (superseded on hard) or declares returned payments.
|
| 30 |
+
- **False premise:** the CFO names three "written-off" loans, and usually 1–2 are not in the register. Follow the records and list those loans as discrepancies. In 25% of instances the CFO is right.
|
| 31 |
+
- **Hard adds:** a future-dated memo, payments recorded after the reporting date, and FX.
|
| 32 |
+
|
| 33 |
+
**Software (3)**
|
| 34 |
+
- **pyrepo-bug1 / implement.** Real bugs in vendored toolz / boltons / tabulate / inflection, graded by the complete pristine test suite.
|
| 35 |
+
- **repair-medium.** A synthetic billing package with SPEC and CHANGELOG, and hidden tests.
|
|
|
|
|
|
|
| 36 |
|
| 37 |
**Office (1)**
|
| 38 |
- **expenses-easy.** A policy audit.
|
| 39 |
|
| 40 |
+
**Industrial (2)**
|
| 41 |
+
- **reflow-hard.** A reflow-oven audit.
|
| 42 |
+
- **downtime-medium.** A sensor-log downtime audit.
|
| 43 |
+
|
| 44 |
+
## Reward
|
| 45 |
+
- **0** if any protected check fails: broken output schema, a changed input file, a test that passed before now failing, an edited test, or the decoy "fixed".
|
| 46 |
+
- Otherwise **full success = 0.9 + 0.1 × steps_left / 12**.
|
| 47 |
+
- Otherwise **partial = 0.5 × share of checks passed**.
|
| 48 |
+
|
| 49 |
+
The reward always lies in [0, 1].
|
| 50 |
|
| 51 |
## Calibration
|
| 52 |
+
Qwen3.8-27B, thinking off, 12,288-token completion budget, 8 episodes per task, counting full solves:
|
| 53 |
+
|
| 54 |
+
| task | strict JSON | lenient JSON |
|
| 55 |
+
|---|---|---|
|
| 56 |
+
| pyrepo-bug1 | 1/8 | 3/8 |
|
| 57 |
+
| pyrepo-implement | 2/8 | 1/8 |
|
| 58 |
+
| repair-medium | 2/8 | 2/8 |
|
| 59 |
+
| loans-easy | 0/8 | 5/8 |
|
| 60 |
+
| loans-medium | 0/8 | 5/8 |
|
| 61 |
+
| loans-hard | 0/8 | 4/8 |
|
| 62 |
+
| expenses-easy | 1/8 | 3/8 |
|
| 63 |
+
| reflow-hard | 3/8 | 0/8 |
|
| 64 |
+
| downtime-medium | 2/8 | 3/8 |
|
| 65 |
+
|
| 66 |
+
Strict failures are almost all format: prose before the JSON, unescaped quotes, `<tool_call>` tags.
|
| 67 |
|
| 68 |
## Licences
|
| 69 |
The vendored libraries keep their own licences: toolz and boltons are BSD-3-Clause, tabulate and inflection are MIT.
|
data/tasks.jsonl
CHANGED
|
@@ -1,10 +1,9 @@
|
|
| 1 |
-
{"task_id": "pyrepo-bug1", "split": "train", "family": "pyrepo", "level": "bug1", "domain": "software-engineering", "checks":
|
| 2 |
-
{"task_id": "pyrepo-
|
| 3 |
-
{"task_id": "
|
| 4 |
-
{"task_id": "
|
| 5 |
-
{"task_id": "
|
| 6 |
-
{"task_id": "
|
| 7 |
-
{"task_id": "
|
| 8 |
-
{"task_id": "
|
| 9 |
-
{"task_id": "downtime-medium", "split": "train", "family": "downtime", "level": "medium", "domain": "industrial
|
| 10 |
-
{"task_id": "expenses-easy", "split": "train", "family": "expenses", "level": "easy", "domain": "office-white-collar", "checks": 8, "qwen3.8_27b_full_score": "5/8", "example_instruction": "# Expense audit\n\nFinance needs an audit of the Research department's expense claims dated in September 2026 (2026-09-01 to 2026-09-30).\n\n- Claims are in /app/data/claims.csv; employees and departments in /app/data/employees.csv.\n- The travel and expense policy is in /app/policy, including amendments. The finance team's folder /app/finance has the daily FX rates and their notes.\n- Apply the policy exactly as written: city tiers, caps, receipt and flight rules, currency conversion and rounding, and how amendments apply.\n\nWrite /app/output/expense_report.json with exactly this structure:\n\n{\n \"violations\": [<IDs of the department's claims in that month that break policy, sorted ascending>],\n \"by_employee\": {\"E01\": <total reimbursable EUR for that employee's claims in scope, 2 decimals>, ...},\n \"reimbursable_eur\": <total reimbursable EUR for all claims in scope, 2 decimals>\n}\n\nInclude in by_employee every employee of the department who has at least one claim in that month.\n", "example_files": ["data/claims.csv", "data/employees.csv", "finance/README.md", "finance/fx_corrections.md", "finance/fx_rates.csv", "policy/amendments/amendment-21_2026-09-02.md", "policy/travel_policy.md"], "n_files": 7}
|
|
|
|
| 1 |
+
{"task_id": "pyrepo-bug1", "split": "train", "family": "pyrepo", "level": "bug1", "domain": "software-engineering", "checks": 5, "protected_checks": ["library imports and the full test suite runs", "every test that passed before still passes", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "1/8", "qwen3_8_27b_full_solves_lenient": "3/8", "example_instruction": "# Fix the tabulate library\n\nA recent change introduced a bug in the `tabulate` library at /app/repo/tabulate. For example, the test `tests.test_output::test_break_long_words` now fails.\n\nFind the cause and fix it in the library code.\n\n- Run the visible tests with `cd /app/repo/tabulate && python3 -m pytest -q tests`. They do not cover everything: hidden tests will run the library's complete test suite.\n- Change only the library's own code under /app/repo/tabulate/tabulate. Do not edit or add tests; edited tests are ignored when grading.\n- Do not break anything that currently works.\n\nCall submit when you are done.\n", "example_files": ["TASK.md", "repo/tabulate/LICENSE", "repo/tabulate/pytest.ini", "repo/tabulate/tabulate/__init__.py", "repo/tabulate/tests/common.py", "repo/tabulate/tests/test_api.py", "repo/tabulate/tests/test_internal.py", "repo/tabulate/tests/test_output.py", "repo/tabulate/tests/test_regression.py"], "n_files": 9}
|
| 2 |
+
{"task_id": "pyrepo-implement", "split": "train", "family": "pyrepo", "level": "implement", "domain": "software-engineering", "checks": 5, "protected_checks": ["library imports and the full test suite runs", "every test that passed before still passes", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "2/8", "qwen3_8_27b_full_solves_lenient": "1/8", "example_instruction": "# Fix the tabulate library\n\nThe function `simple_separated_format` in `/app/repo/tabulate/tabulate/__init__.py` is not implemented yet: it raises NotImplementedError. Implement it so it behaves as its docstring describes, consistent with the rest of the `tabulate` library.\n\n- Run the visible tests with `cd /app/repo/tabulate && python3 -m pytest -q tests`. They do not cover everything: hidden tests will run the library's complete test suite.\n- Change only the library's own code under /app/repo/tabulate/tabulate. Do not edit or add tests; edited tests are ignored when grading.\n- Do not break anything that currently works.\n\nCall submit when you are done.\n", "example_files": ["TASK.md", "repo/tabulate/LICENSE", "repo/tabulate/pytest.ini", "repo/tabulate/tabulate/__init__.py", "repo/tabulate/tests/common.py", "repo/tabulate/tests/test_api.py", "repo/tabulate/tests/test_internal.py", "repo/tabulate/tests/test_output.py", "repo/tabulate/tests/test_regression.py"], "n_files": 9}
|
| 3 |
+
{"task_id": "repair-medium", "split": "train", "family": "repair", "level": "medium", "domain": "software-engineering", "checks": 5, "protected_checks": ["package imports and every function runs", "no regressions in the functions that were already correct", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "2/8", "qwen3_8_27b_full_solves_lenient": "2/8", "example_instruction": "# Fix the accrual package\n\nCustomers report wrong amounts from the invoicing service. The cause is in the Python package at /app/project/src/accrual.\n\n- The required behaviour is in /app/project/docs/SPEC.md and /app/project/docs/CHANGELOG.md (the latest changelog entry wins where they differ).\n- The visible tests in /app/project/tests cover only part of the behaviour. Run them with `cd /app/project && python3 -m pytest -q`.\n- Fix every function that does not match the documented behaviour. Do not change function names, signatures or module layout, and do not break functions that are already correct.\n\nHidden tests will check every function in the package against the documentation. Call submit when you are done.\n", "example_files": ["TASK.md", "project/README.md", "project/docs/CHANGELOG.md", "project/docs/SPEC.md", "project/pytest.ini", "project/scripts/export_invoices.py", "project/src/accrual/__init__.py", "project/src/accrual/dates.py", "project/src/accrual/fees.py", "project/src/accrual/invoices.py", "project/src/accrual/money.py", "project/src/accrual/pricing.py", "project/src/accrual/reports.py", "project/src/accrual/units.py", "project/tests/test_basic.py"], "n_files": 15}
|
| 4 |
+
{"task_id": "loans-easy", "split": "train", "family": "loans", "level": "easy", "domain": "finance", "checks": 7, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "0/8", "qwen3_8_27b_full_solves_lenient": "5/8", "example_instruction": "# Month-end loan provisions\n\nThe CFO writes: \"Please prepare the August 2026 month-end loan-loss provisions as at 2026-08-31. Note that loans L-1002, L-1013 and L-1521 were written off last month, so leave them out.\"\n\n- The loan book is in /app/data/loans.csv, installments and payments in /app/data/payments.csv, and the write-off register in /app/data/writeoffs.csv.\n- The credit policy is /app/policy/credit_policy.md. Risk-committee memos in /app/memos amend it. FX rates are in /app/finance.\n- Apply the policy exactly as written: days past due, stages, provision rates, conversion and rounding, write-offs, and how memos apply.\n- The request may be wrong. Check every claim it makes against the records: where they disagree, follow the records and list the loan IDs concerned under \"discrepancies\". Only loans named in the request can be discrepancies.\n\nWrite /app/output/provisions.json with exactly this structure:\n\n{\n \"stage2_loans\": [<IDs of the loans in stage 2, sorted ascending>],\n \"stage3_loans\": [<IDs of the loans in stage 3, sorted ascending>],\n \"total_provision_eur\": <total provision in EUR, 2 decimals>,\n \"discrepancies\": [<IDs of loans named in the request whose claim the records contradict, sorted ascending; [] if none>]\n}\n", "example_files": ["TASK.md", "data/loans.csv", "data/payments.csv", "data/writeoffs.csv", "finance/fx_rates.csv", "memos/memo-41_2026-06-01.md", "memos/memo-42_2026-06-21.md", "policy/credit_policy.md"], "n_files": 8}
|
| 5 |
+
{"task_id": "loans-medium", "split": "train", "family": "loans", "level": "medium", "domain": "finance", "checks": 7, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "0/8", "qwen3_8_27b_full_solves_lenient": "5/8", "example_instruction": "# Month-end loan provisions\n\nThe CFO writes: \"Please prepare the August 2026 month-end loan-loss provisions as at 2026-08-31. Note that loans L-1128, L-1196 and L-1803 were written off last month, so leave them out.\"\n\n- The loan book is in /app/data/loans.csv, installments and payments in /app/data/payments.csv, and the write-off register in /app/data/writeoffs.csv.\n- The credit policy is /app/policy/credit_policy.md. Risk-committee memos in /app/memos amend it. FX rates are in /app/finance.\n- Apply the policy exactly as written: days past due, stages, provision rates, conversion and rounding, write-offs, and how memos apply.\n- The request may be wrong. Check every claim it makes against the records: where they disagree, follow the records and list the loan IDs concerned under \"discrepancies\". Only loans named in the request can be discrepancies.\n\nWrite /app/output/provisions.json with exactly this structure:\n\n{\n \"stage2_loans\": [<IDs of the loans in stage 2, sorted ascending>],\n \"stage3_loans\": [<IDs of the loans in stage 3, sorted ascending>],\n \"total_provision_eur\": <total provision in EUR, 2 decimals>,\n \"discrepancies\": [<IDs of loans named in the request whose claim the records contradict, sorted ascending; [] if none>]\n}\n", "example_files": ["TASK.md", "data/loans.csv", "data/payments.csv", "data/payments_returned.csv", "data/writeoffs.csv", "finance/fx_rates.csv", "memos/memo-41_2026-06-06.md", "memos/memo-42_2026-06-09.md", "memos/memo-43_2026-08-09.md", "memos/memo-44_2026-08-12.md", "policy/credit_policy.md"], "n_files": 11}
|
| 6 |
+
{"task_id": "loans-hard", "split": "train", "family": "loans", "level": "hard", "domain": "finance", "checks": 7, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "0/8", "qwen3_8_27b_full_solves_lenient": "4/8", "example_instruction": "# Month-end loan provisions\n\nThe CFO writes: \"Please prepare the August 2026 month-end loan-loss provisions as at 2026-08-31. Note that loans L-1006, L-2498 and L-3170 were written off last month, so leave them out.\"\n\n- The loan book is in /app/data/loans.csv, installments and payments in /app/data/payments.csv, and the write-off register in /app/data/writeoffs.csv.\n- The credit policy is /app/policy/credit_policy.md. Risk-committee memos in /app/memos amend it. FX rates are in /app/finance.\n- Apply the policy exactly as written: days past due, stages, provision rates, conversion and rounding, write-offs, and how memos apply.\n- The request may be wrong. Check every claim it makes against the records: where they disagree, follow the records and list the loan IDs concerned under \"discrepancies\". Only loans named in the request can be discrepancies.\n\nWrite /app/output/provisions.json with exactly this structure:\n\n{\n \"stage2_loans\": [<IDs of the loans in stage 2, sorted ascending>],\n \"stage3_loans\": [<IDs of the loans in stage 3, sorted ascending>],\n \"total_provision_eur\": <total provision in EUR, 2 decimals>,\n \"discrepancies\": [<IDs of loans named in the request whose claim the records contradict, sorted ascending; [] if none>]\n}\n", "example_files": ["TASK.md", "data/loans.csv", "data/payments.csv", "data/payments_returned.csv", "data/writeoffs.csv", "finance/fx_rates.csv", "memos/memo-41_2026-06-15.md", "memos/memo-42_2026-07-27.md", "memos/memo-43_2026-07-28.md", "memos/memo-44_2026-08-06.md", "memos/memo-45_2026-08-09.md", "memos/memo-46_2026-08-10.md", "memos/memo-47_2026-08-15.md", "memos/memo-48_2026-08-29.md", "memos/memo-49_2026-09-21.md", "policy/credit_policy.md"], "n_files": 16}
|
| 7 |
+
{"task_id": "expenses-easy", "split": "train", "family": "expenses", "level": "easy", "domain": "office", "checks": 9, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "1/8", "qwen3_8_27b_full_solves_lenient": "3/8", "example_instruction": "# Expense audit\n\nFinance needs an audit of the Research department's expense claims dated in September 2026 (2026-09-01 to 2026-09-30).\n\n- Claims are in /app/data/claims.csv; employees and departments in /app/data/employees.csv.\n- The travel and expense policy is in /app/policy, including amendments. The finance team's folder /app/finance has the daily FX rates and their notes.\n- Apply the policy exactly as written: city tiers, caps, receipt and flight rules, currency conversion and rounding, and how amendments apply.\n\nWrite /app/output/expense_report.json with exactly this structure:\n\n{\n \"violations\": [<IDs of the department's claims in that month that break policy, sorted ascending>],\n \"by_employee\": {\"E01\": <total reimbursable EUR for that employee's claims in scope, 2 decimals>, ...},\n \"reimbursable_eur\": <total reimbursable EUR for all claims in scope, 2 decimals>\n}\n\nInclude in by_employee every employee of the department who has at least one claim in that month.\n", "example_files": ["TASK.md", "data/claims.csv", "data/employees.csv", "finance/README.md", "finance/fx_corrections.md", "finance/fx_rates.csv", "policy/amendments/amendment-21_2026-09-02.md", "policy/travel_policy.md"], "n_files": 8}
|
| 8 |
+
{"task_id": "reflow-hard", "split": "train", "family": "reflow", "level": "hard", "domain": "industrial", "checks": 10, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "3/8", "qwen3_8_27b_full_solves_lenient": "0/8", "example_instruction": "# Reflow profile audit\n\nProcess engineering needs an audit of today's reflow oven runs on line 3.\n\n- The process handbook is in /app/docs (read every document there; definitions and limits come only from the handbook).\n- Thermocouple readings are in /app/data/thermocouples.csv; run details are in /app/data/runs.csv.\n- Use the thermocouple the handbook prescribes for each run. Report temperatures in degrees C.\n\nWrite /app/output/reflow_report.json with exactly this structure:\n\n{\n \"ramp_limit_c_per_s\": <the limit that applies>,\n \"runs\": {\n \"R100\": {\"tc_id\": \"TC2\", \"max_preheat_ramp\": <2 dp>, \"tal_s\": <2 dp>, \"peak_c\": <2 dp>, \"status\": \"compliant\" or \"non-compliant\"},\n ...\n },\n \"failing_runs\": [<run IDs that are non-compliant, sorted>]\n}\nUse null for every value of a run that has no data.\n\nRound to 2 decimals. Do not guess values: derive everything from the files.\n", "example_files": ["TASK.md", "data/boards.csv", "data/runs.csv", "data/thermocouples.csv", "docs/handbook.md", "docs/handbook_revision_B.md"], "n_files": 6}
|
| 9 |
+
{"task_id": "downtime-medium", "split": "train", "family": "downtime", "level": "medium", "domain": "industrial", "checks": 8, "protected_checks": ["output file exists and is a JSON object", "has exactly the required keys with the right types", "input files left unchanged"], "qwen3_8_27b_full_solves_strict": "2/8", "qwen3_8_27b_full_solves_lenient": "3/8", "example_instruction": "# Line downtime report\n\nPlant engineering needs the unplanned downtime of production line A for the evening shift of 2026-09-21.\n\n- The plant documentation is in /app/handbook, including amendments. Use it for the definition of unplanned downtime, the per-shift downtime limit that applies on that date, and any rules about which stops count.\n- Sensor data is in /app/data/sensor_log.csv: one row per machine per minute; `state` is the machine state during that minute.\n- Shift times are in /app/data/shifts.csv.\n\nWrite /app/output/downtime.json with exactly this structure:\n\n{\n \"limit_min\": <the per-shift unplanned downtime limit that applies on 2026-09-21>,\n \"downtime_by_machine\": {\"M01\": <unplanned downtime minutes>, ...},\n \"over_limit\": [<IDs of machines whose downtime is strictly greater than the limit, sorted ascending>]\n}\n\nInclude every machine that appears in the sensor log. Do not guess values: derive everything from the files.\n", "example_files": ["TASK.md", "data/machines.json", "data/sensor_log.csv", "data/shifts.csv", "handbook/amendments/amendment-101_2026-03-10.md", "handbook/amendments/amendment-102_2026-03-25.md", "handbook/amendments/amendment-103_2026-07-16.md", "handbook/amendments/amendment-104_2026-08-16.md", "handbook/maintenance_manual.md"], "n_files": 9}
|
|
|