Workbench v1: 9 tasks, card, calibration

README.md

@@ -0,0 +1,61 @@
+---
+license: bsd-3-clause
+pretty_name: Workbench (OpenEnv)
+tags:
+- openenv
+- reinforcement-learning
+- agents
+- tool-use
+- grpo
+configs:
+- config_name: tasks
+  data_files:
+  - split: train
+    path: data/tasks.jsonl
+---
+
+# Workbench
+
+An OpenEnv environment that trains **workspace habits**:
+- explore a workspace that is too large to read in full
+- search it with grep
+- catch the hidden caveat (an amendment, a correction or an operations note)
+- compute with code
+- verify the result
+- write an exact output file
+
+It was built for [OpenEnv Arena](https://openenvarena-arena.hf.space/).
+
+- **Source:** https://github.com/moarshy/workbench
+- **Image:** `ghcr.io/moarshy/workbench:v1` (linux/amd64)
+
+## Interface
+Each action is one JSON object:
+- `{"type": "run_bash", "command": "..."}` runs as an unprivileged user in `/app`. Output is capped at 8,192 characters, each command has a 120 s timeout, and there is no network. Available tools: python3 + pandas/numpy, grep/rg, jq, sqlite3.
+- `{"type": "submit", "answer": "..."}` ends the task, and the output files are graded.
+
+The answer key is never written to disk, and the server code can't be read from the sandbox. Tasks are graded automatically after 20 actions.
+
+**Reward** = 0.5 × (share of checks passed) + 0.5 × (all checks passed). Each task has 6–8 deterministic checks.
+
+## Tasks (9)
+| task_id | domain | the job | caveats |
+|---|---|---|---|
+| downtime-{easy,medium,hard} | industrial | Per-machine unplanned downtime for one shift, from a per-minute sensor log of up to 52k rows plus a 600-line handbook | micro-stop exclusion, superseded and future-dated amendments, a revised limit, a night shift that crosses midnight |
+| expenses-{easy,medium,hard} | office / finance | Audit a department-month of up to 3,000 multi-currency claims against a travel policy | cap changes mid-month (applied by claim date), superseded caps, a city changing tier, FX corrections in a separate folder, a future-dated rule |
+| routing-{easy,medium,hard} | math (optimisation) | A two-van routing plan with capacity and time windows, within 5% of the optimum | dispatch notes: a forbidden van, a closed road, a cancelled order |
+
+Every reset generates a fresh random workspace. `data/tasks.jsonl` holds one example instruction and file listing per task.
+
+## Calibration
+Qwen3.8-27B, thinking off, 8 episodes per task, counting full-score episodes. The first JSON object in each reply is taken as the action.
+
+| | easy | medium | hard |
+|---|---|---|---|
+| downtime | 7/8 | 5/8 | 4/8 |
+| expenses | 4/8 | 4/8 | 3/8 |
+| routing | 5/8 | 3/8 | 4/8 |
+
+Most failures were reply-format slips: prose before the JSON, unescaped quotes in shell commands, and Qwen's `<tool_call>` format. These are habits that RL can fix.
+
+**Verification:** independent reference solvers that read only the agent-visible files score 1.0 on every tested seed. Solvers that ignore the caveats fail on medium and hard. Sandbox isolation is tested inside the container. See the source repository for details.

data/tasks.jsonl

@@ -0,0 +1,9 @@
+{"task_id": "downtime-easy", "split": "train", "family": "downtime", "level": "easy", "domain": "industrial-physical-systems", "output_file": "output/downtime.json", "checks": 7, "qwen3.8_27b_full_score": "7/8", "example_instruction": "# Line downtime report\n\nPlant engineering needs the unplanned downtime of production line A for the evening shift of 2026-09-20.\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 2026-09-20>,\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": ["data/machines.json", "data/sensor_log.csv", "data/shifts.csv", "handbook/amendments/amendment-101_2026-08-13.md", "handbook/maintenance_manual.md"]}
+{"task_id": "downtime-medium", "split": "train", "family": "downtime", "level": "medium", "domain": "industrial-physical-systems", "output_file": "output/downtime.json", "checks": 7, "qwen3.8_27b_full_score": "5/8", "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": ["data/machines.json", "data/sensor_log.csv", "data/shifts.csv", "handbook/amendments/amendment-101_2026-04-17.md", "handbook/amendments/amendment-102_2026-06-19.md", "handbook/amendments/amendment-103_2026-08-13.md", "handbook/maintenance_manual.md"]}
+{"task_id": "downtime-hard", "split": "train", "family": "downtime", "level": "hard", "domain": "industrial-physical-systems", "output_file": "output/downtime.json", "checks": 7, "qwen3.8_27b_full_score": "4/8", "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": ["data/machines.json", "data/sensor_log.csv", "data/shifts.csv", "handbook/amendments/amendment-101_2026-02-24.md", "handbook/amendments/amendment-102_2026-02-28.md", "handbook/amendments/amendment-103_2026-03-10.md", "handbook/amendments/amendment-104_2026-03-16.md", "handbook/amendments/amendment-105_2026-06-17.md", "handbook/amendments/amendment-106_2026-08-16.md", "handbook/amendments/amendment-107_2026-11-05.md", "handbook/maintenance_manual.md"]}
+{"task_id": "expenses-easy", "split": "train", "family": "expenses", "level": "easy", "domain": "office-white-collar / finance-economics", "output_file": "output/expense_report.json", "checks": 8, "qwen3.8_27b_full_score": "4/8", "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"]}
+{"task_id": "expenses-medium", "split": "train", "family": "expenses", "level": "medium", "domain": "office-white-collar / finance-economics", "output_file": "output/expense_report.json", "checks": 8, "qwen3.8_27b_full_score": "4/8", "example_instruction": "# Expense audit\n\nFinance needs an audit of the Finance 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_2026-08-15.md", "policy/amendments/amendment-22_2026-09-13.md", "policy/amendments/amendment-23_2026-09-15.md", "policy/travel_policy.md"]}
+{"task_id": "expenses-hard", "split": "train", "family": "expenses", "level": "hard", "domain": "office-white-collar / finance-economics", "output_file": "output/expense_report.json", "checks": 8, "qwen3.8_27b_full_score": "3/8", "example_instruction": "# Expense audit\n\nFinance needs an audit of the Finance 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_2026-07-13.md", "policy/amendments/amendment-22_2026-07-18.md", "policy/amendments/amendment-23_2026-09-14.md", "policy/amendments/amendment-24_2026-09-16.md", "policy/amendments/amendment-25_2026-09-18.md", "policy/amendments/amendment-26_2026-09-21.md", "policy/amendments/amendment-27_2026-11-19.md", "policy/travel_policy.md"]}
+{"task_id": "routing-easy", "split": "train", "family": "routing", "level": "easy", "domain": "mathematics-or-formal-reasoning", "output_file": "output/plan.json", "checks": 6, "qwen3.8_27b_full_score": "5/8", "example_instruction": "# Depot delivery plan\n\nPlan today's deliveries for two vans, V1 and V2.\n\n- Orders are in /app/data/orders.csv; the depot, vans, speed and service time are in /app/data/vans.json.\n- Read /app/data/ops_notes.md: every note there is a binding constraint.\n- Distances are great-circle distances on a sphere of radius 6371 km between the coordinates. Travel time = distance / speed; each stop takes the service time.\n- Respect van capacity, and every note. Each remaining order is served exactly once.\n- Minimise the total distance driven by both vans, including the return to the depot.\n\nWrite /app/output/plan.json with exactly this structure:\n\n{\n  \"routes\": {\"V1\": [<order IDs in visiting order>], \"V2\": [<order IDs in visiting order>]},\n  \"total_km\": <total distance of your plan in km, 2 decimals>\n}\n\nA plan is accepted if it is feasible and its total distance is within 5% of the best possible plan.\n", "example_files": ["data/ops_notes.md", "data/orders.csv", "data/vans.json"]}
+{"task_id": "routing-medium", "split": "train", "family": "routing", "level": "medium", "domain": "mathematics-or-formal-reasoning", "output_file": "output/plan.json", "checks": 6, "qwen3.8_27b_full_score": "3/8", "example_instruction": "# Depot delivery plan\n\nPlan today's deliveries for two vans, V1 and V2.\n\n- Orders are in /app/data/orders.csv; the depot, vans, speed and service time are in /app/data/vans.json.\n- Read /app/data/ops_notes.md: every note there is a binding constraint.\n- Distances are great-circle distances on a sphere of radius 6371 km between the coordinates. Travel time = distance / speed; each stop takes the service time.\n- Respect van capacity, time windows (a van may wait for a window to open), and every note. Each remaining order is served exactly once.\n- Minimise the total distance driven by both vans, including the return to the depot.\n\nWrite /app/output/plan.json with exactly this structure:\n\n{\n  \"routes\": {\"V1\": [<order IDs in visiting order>], \"V2\": [<order IDs in visiting order>]},\n  \"total_km\": <total distance of your plan in km, 2 decimals>\n}\n\nA plan is accepted if it is feasible and its total distance is within 5% of the best possible plan.\n", "example_files": ["data/ops_notes.md", "data/orders.csv", "data/vans.json"]}
+{"task_id": "routing-hard", "split": "train", "family": "routing", "level": "hard", "domain": "mathematics-or-formal-reasoning", "output_file": "output/plan.json", "checks": 6, "qwen3.8_27b_full_score": "4/8", "example_instruction": "# Depot delivery plan\n\nPlan today's deliveries for two vans, V1 and V2.\n\n- Orders are in /app/data/orders.csv; the depot, vans, speed and service time are in /app/data/vans.json.\n- Read /app/data/ops_notes.md: every note there is a binding constraint.\n- Distances are great-circle distances on a sphere of radius 6371 km between the coordinates. Travel time = distance / speed; each stop takes the service time.\n- Respect van capacity, time windows (a van may wait for a window to open), and every note. Each remaining order is served exactly once.\n- Minimise the total distance driven by both vans, including the return to the depot.\n\nWrite /app/output/plan.json with exactly this structure:\n\n{\n  \"routes\": {\"V1\": [<order IDs in visiting order>], \"V2\": [<order IDs in visiting order>]},\n  \"total_km\": <total distance of your plan in km, 2 decimals>\n}\n\nA plan is accepted if it is feasible and its total distance is within 5% of the best possible plan.\n", "example_files": ["data/ops_notes.md", "data/orders.csv", "data/vans.json"]}