Syndicate604/science-desk

Science Desk

An OpenEnv environment for the OpenEnv Arena: natural-science data tasks in the style of public SkillsBench v1.1 science tasks, served through the same tbench2_env interface the arena uses for its own tasks. Each episode is a workspace under /app with a short instruction. The agent runs commands (exec), writes files (write_file) and sends evaluate; a hidden verifier checks the one output file the task names and returns 1 or 0.

Image: ghcr.io/syndicate604/science-desk (linux/amd64). Source: source/ in this repo. Forked from dnakhla/terminal-desk v2 (Apache-2.0); see NOTICE.

Interface

Terminal Desk v2's copy of OpenEnv's upstream tbench2_env contract: exec, write_file, evaluate, close; the instruction repeated in every observation; output not cut; per-task budgets of 4096 completion tokens (model tokens plus every observation after the first), 8192 context tokens and 32 tool calls. The image has numpy, scipy, pandas and astropy.

Tasks

10 families x 3 levels = 30 task ids (<family>-L<level>); each id is one fixed problem, so a GRPO group compares answers to the same question. <family>-L<level>-i<k> names further fixed instances. Eight families are modelled on SkillsBench science tasks with generated data: sci_units (lab-unit-harmonization), sci_protein (protein-expression-analysis), sci_flood (flood-risk-analysis), sci_trend (lake-warming-attribution), sci_quake (earthquake-plate-calculation), sci_period (exoplanet-detection-period), sci_picking (seismic-phase-picking), sci_vitals (radar-vital-signs). sci_kinetics and sci_stoich come from Terminal Desk. tasks.jsonl lists every id with its fixed instruction, output file and screen numbers.

Every task is checked by tests: the reference answer scores 1, missing or corrupted output scores 0, and an independent solver that reads only the files and the instruction scores 1 on four instances per level.

Submission (v5)

selected_tasks.txt, 12 tasks with both successes and failures in a screen of Qwen3.8-27B (thinking off) via OpenRouter, DeepInfra bf16 pinned, 6 episodes each, under the budgets above with observation tokens counted by the Qwen tokenizer on arrival:

Task	Solved / 6
sci_picking-L1 / L3	3 / 1
sci_units-L2 / L3	3 / 3
sci_quake-L3	2
sci_trend-L3	4
sci_flood-L2, sci_protein-L3, sci_stoich-L3	5 each
sci_period-L1 / L2 / L3 (with astropy)	3 / 3 / 2

Failures include malformed action replies, stopping before writing the output, budget exhaustion and wrong results. A malformed final reply ends the episode through evaluate, so a correct file already written still scores 1.

Limitations: six episodes per task is noisy, and earlier screens underestimated the arena-trained model, so these numbers are not predictions of arena solve rates. Twelve ids are twelve fixed problems: this is an interface-and-workflow experiment, not broad science coverage.

Build and test

See source/README.md.