NoeFlandre/workbench-gym

Workbench Desk v7

Workbench Desk trains file inspection, numerical analysis, and production of verified output files through shell and file tools. It contains 34 task IDs, with two difficulty levels for each of 17 procedural families across the Arena's eight domains. Each reset creates fresh data from a server-selected random seed. Training uses the public container image, which contains all generators, assets, and grading code. It downloads nothing at startup.

The task manifest is tasks_v7.jsonl. The complete environment source and reference solvers are under releases/v7/source/, with the pinned OpenEnv wheel. Historical task manifests remain available in this dataset.

Actions and rewards

The tools are bash, write, and submit. A task starts with a prose instruction and files such as CSV, XLSX, JSON, and logs. The instruction names the required output files. The agent has 20 calls, and submit ends the episode. The finish action is {"tool":"submit","args":{}}.

An exact answer earns 1. Otherwise, reward equals 0.1 times the fraction of checked values in a valid format, plus 0.1 times the fraction of correct values, capped at 0.2. Missing output files earn 0. Invalid values never earn full reward. The grader reads bounded regular files and rejects symbolic links. Agent shell commands run as an unprivileged user with CPU, memory, output, and wall-clock limits. The grader is inaccessible to that user.

Level 2 adds data interpretation issues such as revisions, status rows, units, sentinel values, and corrections in notes. Task families include finance, science, math, media, office analysis, industrial systems, software, and security. The source includes independent solvers that derive results from input files and instructions, plus plausible wrong-answer controls.

Verification and limits

The release is checked through an anonymous pull of its public amd64 image, the pinned OpenEnv validator, and 204 fresh WebSocket episodes. Each task has exact-answer, empty, malformed, wrong-answer, example, and finish controls. Reference solvers operate on generated input files. No LLM inference or model calibration is used to select this release. These checks establish environment behavior; they do not predict improvement on private Arena tasks.

The release manifest links the exact image digest, source commit, and public verification run. A prior v6 release scored 3/40 on the Arena evaluation. No v7 training or evaluation result exists before an approved submission.