Syndicate604/skills-desk-v4

Skills Desk

An OpenEnv environment for the OpenEnv Arena that trains reliable problem solving with tools: understand the request, find the relevant information, compute or repair something, produce the required output, and check it. Each episode is a small workspace under /app with a short instruction. The agent runs commands (exec), writes files (write_file) and sends evaluate; a hidden grader checks the result and returns 1 or 0.

Image: ghcr.io/syndicate604/skills-desk (linux/amd64). Source: source/. Forked from Science Desk and 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 Python 3.12 with numpy, scipy, pandas and astropy, plus jq, bc and sqlite3.

Training tasks (v4: 23)

manifest/training_tasks_v4.txt. A GRPO step updates the model only when its group of rollouts is mixed (some solved, some not); in the previous two Skills Desk runs only 21-28% of steps were. v4 trains only on tasks with evidence of mixed groups (manifest/v4-MIX.md, per-task numbers in manifest/v4-mix-selection.json):

• Arena evidence: tasks visited at least 4 times in earlier Syndicate604 arena runs (Skills Desk v1-r3 and v3, Science Desk v5) with mixed groups on at least 40% of visits. 12 Skills tasks and 6 Science Desk tasks; together they were mixed on 60% of past visits.
• Screen evidence: 5 new, harder output-format tasks whose content solve rate under the arena's budgets makes a mixed group of 4 likely.

This selection was made from observed training logs, not pre-registered. Each id is one fixed problem; ids shared with earlier releases serve exactly their earlier instances.

Source	Tasks
Skills (evidence, compute, transform, repair)	events log to JSON, Markdown table to JSON, deduplication script, backorders L2, calibration windows, final grades, display numbers L3 (German style)
Skills (constraints)	shift roster L1, course selection, fewest disks, delivery days, correct a one-error roster
Science Desk v5 (natural science)	seismic phase picks L1 and L3, lab unit harmonisation L3, reaction stoichiometry L3, earthquake catalogue L3, climate trend L3
Output format, hard (new)	Berlin-time timeline across DST with ISO durations (L2, L3); CSV after ordered change requests (L1-L3)

Science Desk's families (sci_*) are served unchanged, with their own instance salt and their own lenient CSV reading, so each id serves exactly the instance Science Desk v5 trained on. The new format families fmt_money (currencies and locale styles) and fmt_measure (value plus or minus uncertainty by the Particle Data Group rule) are served but not trained: the screen found no mixed outcomes for them.

Reserved evaluation tasks (20)

manifest/reserved_tasks.txt: four per skill, two fresh applications of a familiar skill and two combining skills, with different file layouts and solution structures. They are served by the same image so a trained checkpoint can be evaluated on them, but they are never trained on and never used to choose or tune the training tasks. The set is frozen (v1); a change would be published as a new version.

tasks.jsonl lists the 23 training and 20 reserved ids with their split, skill, output path and fixed instruction.

Reward

Binary on evaluate. Content and structure are graded together: correct JSON with wrong values fails, correct values in a malformed file fail, and a CSV row with extra or missing fields fails. Harmless differences pass (key order, CSV quoting, line endings, and either rounding of an exact half where the task states no rounding rule; a task that states half-up rounding requires it). Tasks with many correct answers (rosters, schedules, plans) are graded against their rules. Repair tasks run the repaired code as nobody on hidden input. Nothing rewards a particular tool sequence, reply length or claims of checking.

Verification

Every task: the reference output scores 1, no output and corrupted output score 0, an independent agent-style reference script scores 1 on four instances inside the image, and a deliberate common mistake scores 0. Instances are byte-identical in any process and frozen by fingerprints in source/tests/fixtures/. See source/README.md to build and test.

Limitations

The selection screen is small (6 episodes per task) and uses the untrained model through a provider, not the arena's sampler; earlier screens underestimated the arena-trained model, and no correction was applied. Whether training on these tasks transfers can only be measured by evaluating a trained checkpoint on the reserved tasks; the arena's private score is a separate measure.