rca-umb/five-desks-siblings

Five Desks v8: SkillsBench siblings, re-tuned on v7's arena rewards

An OpenEnv environment for the OpenEnv Arena. Every task is a sibling of a named public SkillsBench v1.1 task. It does the same kind of work, uses the same input formats and tools, gives the same style of instruction, and has the same output contract and verifier strictness. Its data and question are new and vary by seed. No public task is included.

Why

• The likely eval. The arena's eight domain ids are SkillsBench v1.1's task taxonomy, verbatim, and the operator publishes both SkillsBench and the arena's Terminal-Bench 2 smoke test (calibration/eval-source.md). So the private tasks are probably SkillsBench-style: real files, a prose brief, terminal and Python, and an output file checked by tests.
• The bar. Swapping public tasks for ours should not change what a trained model learns.
  • Our previous candidate (SkillsBench-flavoured tasks written from scratch) fell short of that bar. Its inputs were 40x smaller and its solutions 6x shorter, with narrower formats and a lenient checker (calibration/v7-terminal-vs-skillsbench.md).
  • This set was built task by task from the public packages (instructions, Dockerfiles, verifiers), read as data.

Tasks: 29 families, levels 0-2 (63 task ids in the image, 22 submitted)

domain	sibling family	public task it extends
finance-economics	finance-invoice-screening	invoice-fraud-detection
finance-economics	finance-xlsx-recover	xlsx-recover-data
finance-economics	finance-hp-correlation	econ-detrending-correlation
finance-economics	finance-commission-qa	financial-modeling-qa
industrial-physical-systems	industrial-stl-mass	3d-scan-calc
industrial-physical-systems	industrial-grid-dispatch	grid-dispatch-operator
industrial-physical-systems	industrial-thermal-control	hvac-control
industrial-physical-systems	industrial-schedule-repair	manufacturing-fjsp-optimization
natural-science	science-snowpack-attribution	lake-warming-attribution
natural-science	science-flood-days	flood-risk-analysis
natural-science	science-transit-period	exoplanet-detection-period
natural-science	science-lab-harmonize	lab-unit-harmonization
office-white-collar	office-docx-template	offer-letter-generator
office-white-collar	office-pdf-form	court-form-filling
office-white-collar	office-pptx-sources	pptx-reference-formatting
software-engineering	software-prefix-cache	llm-prefix-cache-replay
software-engineering	software-fix-build	fix-build-agentops
software-engineering	software-dialogue-graph	dialogue-parser
software-engineering	software-bgp-audit	azure-bgp-oscillation-route-leak
mathematics-or-formal-reasoning	math-pddl-logistics	pddl-tpp-planning
mathematics-or-formal-reasoning	math-bike-rebalance	bike-rebalance
mathematics-or-formal-reasoning	math-trip-planner	travel-planning
mathematics-or-formal-reasoning	math-hex-districts	civ6-adjacency-optimizer
media-content-production	media-dead-air	video-silence-remover
media-content-production	media-threejs-obj	threejs-to-obj
media-content-production	media-sprite-count	mario-coin-counting
media-content-production	media-dub-fit	multilingual-video-dubbing
cybersecurity	security-pcap-stats	dapt-intrusion-detection
cybersecurity	security-lockfile-audit	software-dependency-audit

Each family's module docstring (source/termdesk/siblings/) says what differs from its public task.

Which task ids train (v8)

v8 uses v7's run as feedback (calibration/run-v7-result.md, per task in calibration/run-v7-per-task.json). v7 scored 1/40. Its training moved, though: reward rose from 0.31 to 0.46 over 145 steps, with groups of 4 and one task per step. The arena's own first-pass rewards replace our outside calibration, which was too pessimistic.

• Dropped:
  • tasks that saturated, solved 4/4 on most visits: snowpack-l1, prefix-cache-l1, xlsx-recover-l1, flood-l1;
  • dead tasks outside the target domains: invoice, PDF form, grid dispatch, fix-build, commission, dubbing, dead-air, thermal control.
• Harder level for saturated tasks: snowpack-l2, stl-mass-l2, xlsx-recover-l2, hex-districts-l2, pcap-stats-l2.
• New easier level 0 for dead tasks in the target domains: transit-period-l0, lab-harmonize-l0, pddl-logistics-l0, trip-planner-l0, lockfile-audit-l0. Each is a smaller version of the same work.
• Kept: tasks with mixed outcomes, 4-5 visits each, in the 0.2-0.8 band.
• Weighting: 12 of the 22 submitted ids are in natural science, math and cybersecurity, where a single private solve would add a point to the leaderboard. Each id gets about 6-7 visits in a run of about 145 steps.

Two levels

• Level 2 (-l2) is the public task's shape.
• Level 1 (-l1) is a smaller version of the same work: fewer pages or rows, fewer rules, or one part of the analysis.

The levels exist because of the arena's budget. Under the default limits, the untrained model solves level-2 tasks very rarely (see Calibration). An RL group that never succeeds teaches nothing.

Changes for the arena's limits

• Shrink, don't drop. Public tasks that cannot fit 4,096 / 8,192 tokens are scaled down in quantity: five planning problems instead of thirty, a few minutes of video, a smaller network. The kind of work stays the same.
• Local data. Data the public tasks download (NWS and USGS flood data) is shipped in the agencies' own formats.
• One image. It carries the union of the tools the public tasks install.

Contract

• Schema. The Terminal-Bench 2 schema of OpenEnv's tbench2_env at revision 86a180e, as in the operator's smoke test: exec, write_file, evaluate. Every observation repeats the instruction.
• Workspace. Commands run as nobody in /root, which holds the task's files. The checker and expected answers are out of reach.
• Limits per task. 4,096 completion / 8,192 context tokens, 32 tool calls, 110 s per tool call.
• Reward. 1.0 when every check passes, otherwise 0. No partial credit.

Tools in the image:

• System: python 3.12, ffmpeg, poppler, node with three.js (/root/node_modules), suricata (not used by the submitted tasks).
• Data and science: numpy, pandas, scipy, scikit-learn, statsmodels, pymannkendall, astropy.
• Documents: openpyxl, xlrd, pypdf, pdfplumber, pymupdf, reportlab, python-docx, python-pptx.
• Other: rapidfuzz, pyperplan, PuLP, networkx, scapy, opencv-headless, pytest.

Checks

• Tests in the image (source/tests/test_siblings.py). For every task id:
  • the oracle solves seeds 0, 1 and 7 within the default budget;
  • a plausible wrong answer scores 0;
  • doing nothing scores 0;
  • instances vary by seed.
• Stress runs. Oracle and wrong answer on 12-40 seeds per task id; builds on 100-150 random seeds (media: 12-15). 0 failures after fixes.
• Replay. Over the WebSocket API of the pushed image: 315 episodes (example action, oracle and wrong answer on two seeds per task, all 63 ids), 0 failures (calibration/replay.json).
• Restricted runtime. Without CAP_SETUID/SETGID/CHOWN, resets and evaluate still work, and commands return an error observation.

Calibration (untrained Qwen3.8-27B)

calibration/calibration-summary.md; harness source/scripts/calibrate_tb2.py.

• Setup. qwen-3.8-27b on Cerebras, thinking off, temperature 1.0. Arena rules: exactly one JSON object per reply, the default budgets, 8 episodes per task id.
• Result. 28 solves in 464 episodes (6.0%), on 13 of the 58 task ids. Most episodes ended before any answer was graded:
  • not one JSON object: 278
  • out of budget: 100
  • reply cut at the room left: 66
• Why replies fail. Most malformed replies are valid JSON wrapped in <tool_call> tags or preceded by prose. This matches dnakhla's measurement on the public tasks under the same contract (board #202).
• What training can fix. That reply-format and economy habit, not domain knowledge.
• Correction after v7. The arena's first-pass rewards were far above these numbers on most tasks (for example 1.00 against 1/8). The arena's model breaks the reply format much less often than Qwen did on Cerebras, so this calibration overstates the format problem. v8 selects tasks on the arena's own rewards instead.

Pre-registration (five-desks-v8)

Written before submission.

• Training reward. Starts near 0.5, because the selected tasks sat around 0.2-0.8 in v7, and should rise. At least 60% of steps should carry a gradient; below that, read nothing from the score.
• Private score. Most likely 1-4 of 40, the same band as v7 and the field. Read it against the no-update runs (0-2).
• What would count: a solve in natural science, math or cybersecurity (one run in about 80 each), or a total of 5 or more.

Limits

• Inputs are smaller than the public ones. Instruction length matches: level-2 median 1,112 characters (IQR 841-1,545) against the public 1,212 (811-1,806). Input files are smaller: level-2 median 14 KB (IQR 4 KB-200 KB) against the public 143 KB (18 KB-2.4 MB). This is partly deliberate (shrunk to fit the token budget) and partly because the public set includes long videos and large raw datasets.
• Swap-equivalence is approximated, not proven. The siblings were designed from the public packages, and the checks above confirm they are solvable and strict. A direct comparison covers only a sample (calibration/swap-check.md):
  • The untrained model played four public tasks in their own Docker images, with their own verifiers (lake-warming, 3d-scan, prefix-cache replay, invoice fraud), and the matching level-2 siblings, under the same rules.
  • Public tasks: 0 of 32 solved. Siblings: 0 of 32 solved. Failure mixes were similar: mostly malformed replies, then budget and cut-off replies.
  • That is the right starting point, but it is a small sample and says nothing about a trained model.
• The eval is unconfirmed. Its harness, budgets and tasks are inferred from public data; the organizers have not confirmed them.
• Uneven coverage. Six domains have four families; office has three and security two.

Source

source/:

• termdesk/ holds the server and the families: siblings/ here, plus the older terminal families that are not submitted.
• tests/test_siblings.py, scripts/calibrate_tb2.py and scripts/replay_siblings.py.
• Image: ghcr.io/rca-umb/five-desks:siblings-cand3. The exact request is submission_request.json.