Syndicate604/skills-desk-v5

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 (v5: 35, answer format only)

manifest/training_tasks_v5.txt. v5 trains one skill: delivering any answer in exactly the requested form. Skills Desk v3 and v4 (formatting at 8 of 48 and 5 of 23 tasks) scored 4/40, the same as the untrained model. Every v5 task gives correct facts and a written answer contract; only the form of the answer is graded, all or nothing, character by character or by strict parse.

• JSON answers (fmt_json, L1 and L3): exact key names and order, integer vs number vs string, null for unknown values, booleans from yes/Y/✓, half-up rounding at ties, ISO dates from slash and spelled dates, enum casing, locale numbers, percentages as fractions or basis points, arrays sorted with a tie-break; L3 ends with an amendment that renames a key and changes rounding and units.
• Tables (fmt_table, L1-L3): exact header and column order, CSV quoting only when needed, per-column decimals, units stripped or converted, a missing-value token, date formats, a filter, a multi-key sort and a TOTAL row from rounded values; L3 as a Markdown table or TSV, with an amendment.
• Final-answer files (fmt_text, L1 and L3): <id>: <answer> lines in numeric id order, integers, money with separators and currency position, percentages, fractions in lowest terms, booleans, cleaned text; L3 adds significant figures, scientific notation, unit conversion, durations, dates, placeholders for missing answers, ~ for approximate answers and an amendment.
• Each problem's topic is drawn from eight areas (security, finance, industrial, mathematics, media, natural science, office, software). Instances i0-i3 of a level are separate ids: different topic and data, same rules.
• Also: fmt_money L2-L3 (currencies, locale styles, minor units, conversion and totals) and five format tasks that were mixed in the v4 run (fmt_times L2-L3, fmt_changes L2-L3, fmt_numbers-L3).

Levels were kept when a capped screen under the arena's budgets made a mixed group of 4 likely (manifest/v5-FORMAT.md, numbers in manifest/v5-format-selection.json). Each id is one fixed problem; ids shared with v4 serve exactly their v4 instances.

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 35 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.