| --- |
| license: bsd-3-clause |
| pretty_name: Workbench (OpenEnv) |
| tags: |
| - openenv |
| - reinforcement-learning |
| - agents |
| - tool-use |
| - software-engineering |
| - finance |
| configs: |
| - config_name: tasks |
| data_files: |
| - split: train |
| path: data/tasks.jsonl |
| --- |
| |
| # Workbench (v4 task set: finance traps + protected reward) |
|
|
| An OpenEnv environment that trains **workspace habits** in a `run_bash` / `submit` sandbox, built for [OpenEnv Arena](https://openenvarena-arena.hf.space/). |
|
|
| - **Source:** https://github.com/moarshy/workbench |
| - **Image:** `ghcr.io/moarshy/workbench:v4` |
| - **Notes and pre-registration:** https://github.com/moarshy/openenv-arena-notes/blob/main/versions/v8-workbench-v4.md |
|
|
| ## Tasks (9) |
| **Finance (3), new in v4** |
| - **loans-easy / medium / hard.** Month-end loan-loss provisions: stage loans by days past due, apply provision rates, write `provisions.json`. |
| - **Late caveat:** a dated risk-committee memo changes a rate (superseded on hard) or declares returned payments. |
| - **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. |
| - **Hard adds:** a future-dated memo, payments recorded after the reporting date, and FX. |
|
|
| **Software (3)** |
| - **pyrepo-bug1 / implement.** Real bugs in vendored toolz / boltons / tabulate / inflection, graded by the complete pristine test suite. |
| - **repair-medium.** A synthetic billing package with SPEC and CHANGELOG, and hidden tests. |
|
|
| **Office (1)** |
| - **expenses-easy.** A policy audit. |
|
|
| **Industrial (2)** |
| - **reflow-hard.** A reflow-oven audit. |
| - **downtime-medium.** A sensor-log downtime audit. |
|
|
| ## Reward |
| - **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". |
| - Otherwise **full success = 0.9 + 0.1 × steps_left / 12**. |
| - Otherwise **partial = 0.5 × share of checks passed**. |
| |
| The reward always lies in [0, 1]. |
| |
| ## Calibration |
| Qwen3.8-27B, thinking off, 12,288-token completion budget, 8 episodes per task, counting full solves: |
| |
| | task | strict JSON | lenient JSON | |
| |---|---|---| |
| | pyrepo-bug1 | 1/8 | 3/8 | |
| | pyrepo-implement | 2/8 | 1/8 | |
| | repair-medium | 2/8 | 2/8 | |
| | loans-easy | 0/8 | 5/8 | |
| | loans-medium | 0/8 | 5/8 | |
| | loans-hard | 0/8 | 4/8 | |
| | expenses-easy | 1/8 | 3/8 | |
| | reflow-hard | 3/8 | 0/8 | |
| | downtime-medium | 2/8 | 3/8 | |
| |
| Strict failures are almost all format: prose before the JSON, unescaped quotes, `<tool_call>` tags. |
| |
| ## Licences |
| The vendored libraries keep their own licences: toolz and boltons are BSD-3-Clause, tabulate and inflection are MIT. |
| |
