task_id stringclasses 8
values | image stringclasses 1
value | submitted_for_training bool 1
class | family stringclasses 8
values | domain stringclasses 8
values | max_tool_calls int64 8 16 | example_prompt stringclasses 8
values | untrained_calls_per_episode listlengths 24 28 | untrained_control_reward float64 0.19 0.64 |
|---|---|---|---|---|---|---|---|---|
early-software-repofix-l2 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | software-repofix | software-engineering | 12 | SOFTWARE. A small Python repository has an open bug report. Fix the bug so the release check passes, then submit. Files: ISSUE.md, README.md, issue_comments.md, pager.py, test_pager.py.
Reply with exactly one JSON object per turn: {"tool": "<name>", "args": {...}}. Tools:
- files {}: list the workspace files
- show_fi... | [
7,
7,
8,
6,
7,
6,
7,
6,
8,
6,
7,
7,
9,
7,
7,
7,
6,
6,
7,
6,
8,
9,
9,
7,
7,
8,
9,
7
] | 0.485 |
early-industrial-telemetry-l3 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | industrial-telemetry | industrial-physical-systems | 12 | INDUSTRIAL. Control room, cooling-water loop. An operator heard noise from the pumps. Using the manual and the minute-by-minute telemetry (minutes 0-179), find the first pump to start cavitating this shift and the minute cavitation began. Submit {"answer": {"asset_id": "P-...", "cavitation_start_minute": <int>}}.
Repl... | [
9,
11,
7,
5,
7,
8,
12,
10,
6,
7,
9,
7,
7,
7,
6,
8,
9,
8,
9,
8,
8,
6,
8,
8
] | 0.424 |
early-science-lab-l3 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | science-lab | natural-science | 10 | SCIENCE. Analyse the experiment in this workspace (protocol.md, data.csv, notebook.md) as the protocol describes. Submit {"answer": {"doubling_time_min": <number>, "rows_used": <integer>}}; numbers are checked to 1%.
Reply with exactly one JSON object per turn: {"tool": "<name>", "args": {...}}. Tools:
- list_files {}... | [
6,
7,
5,
5,
6,
6,
6,
5,
6,
6,
6,
5,
6,
6,
6,
6,
5,
5,
6,
6,
6,
6,
6,
6
] | 0.521 |
early-office-scheduling-l1 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | office-scheduling | office-white-collar | 14 | OFFICE. You are an executive assistant. Handle the scheduling request in the inbox: find the right slot and send the invitation with create_event.
Reply with exactly one JSON object per turn: {"tool": "<name>", "args": {...}}. Tools:
- inbox {}: list the inbox
- get_mail {"id": string}: read one email
- busy_times {"w... | [
4,
6,
4,
8,
5,
5,
5,
7,
8,
5,
9,
8,
8,
4,
11,
5,
7,
5,
6,
4,
5,
5,
11,
8
] | 0.616 |
early-finance-receivables-l2 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | finance-receivables | finance-economics | 12 | FINANCE. Accounts receivable database (SQLite, read-only) with tables customers, invoices, payments, credit_notes, refunds and fx_rates. Find duplicate payments: a payment is a duplicate when the same invoice already received a payment of exactly the same amount within the previous 3 days. How many duplicates are still... | [
12,
12,
11,
9,
11,
9,
9,
10,
11,
8,
11,
11,
11,
12,
11,
11,
12,
12,
10,
12,
10,
12,
12,
8
] | 0.191 |
early-math-problems-l2 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | math-problems | mathematics-or-formal-reasoning | 8 | MATH. Solve the problem in problem.md (workspace files: problem.md, clarifications.md). Submit {"answer": {"value": "<your answer>"}}.
Reply with exactly one JSON object per turn: {"tool": "<name>", "args": {...}}. Tools:
- files {}: list the workspace files
- read_file {"path": string}: show one file
- python {"code"... | [
5,
2,
3,
4,
3,
4,
3,
5,
4,
3,
4,
4,
5,
4,
4,
4,
4,
3,
3,
5,
3,
6,
3,
5
] | 0.641 |
early-security-audit-l3 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | security-audit | cybersecurity | 12 | SECURITY. Review this Python web service and report the one vulnerability in the service as it runs in production. Submit {"finding": {"file": <path>, "line": <1-based line>, "cwe": <id>}} using one of: CWE-89 SQL injection; CWE-22 path traversal; CWE-78 OS command injection; CWE-798 hard-coded credentials; CWE-862 mis... | [
10,
10,
8,
10,
10,
2,
10,
10,
4,
9,
4,
11,
3,
9,
10,
11,
9,
2,
10,
2,
4,
10,
9,
7
] | 0.444 |
early-media-copyedit-l3 | ghcr.io/rca-umb/five-desks@sha256:8d805ada53196f55dded455daec7d1a96347dcdf4458b875d2b41ee7a35f9471 | true | media-copyedit | media-content-production | 16 | MEDIA. Copy desk. Edit the news item so it follows the house style guide, then submit.
1: Bluefield Opens New Repair Hub In Tacoma
2: Every visitor gets a free Fix Hub starter pack, which is very popular.
3: Residents can drop off items starting Monday!
4: Visits rose 3 percent in the pilot, with 20 volunteers trained... | [
12,
14,
13,
14,
14,
16,
13,
12,
12,
16,
16,
16,
13,
12,
16,
13,
13,
16,
13,
16,
12,
16,
16,
13
] | 0.185 |
Five Desks v5: a collapse test (positive control)
This is a deliberate diagnostic, not an attempt to improve the model. It asks one question: does training in the arena reach the private-task score at all?
naklitechie's null control (no gradient at all) scored 1/40, and across the 45 evaluated runs so far, very different
training runs have produced statistically indistinguishable private scores (see
calibration/lottery-analysis.md). Two explanations fit: training changes the
model too little to matter on unrelated tasks, or the scored model is not the trained one. A run that trains a
clearly destructive habit tells them apart.
What it trains
The tasks are Five Desks' normal eight-domain tool tasks with their normal briefs (early-<task id>; the model is
not told anything unusual). The reward ignores the answer and pays only for ending the episode early: 1.0 when the
first call ends it, falling linearly to 0 when the tool budget runs out. The untrained model's episode lengths vary
on every task, so every GRPO group has a gradient toward stopping sooner, until it submits or finishes immediately.
Pre-registered prediction
- If training reaches the score: the trained model stops investigating, and the private score falls to 0 or 1 out of 40 (no run has scored 0 so far; about 9% of runs score 1 or less).
- If the score is 2 or more while the training curve shows near-immediate endings, the private evaluation is not measuring the trained policy (or the habit does not carry over to the evaluation's tasks), and we will report it to the operators with the training curve.
Run it yourself
The image is public and the exact request is submission_request.json. Anyone can
submit it from their own account (change submission_id, and dataset if you want the run attributed to your own
dataset). Independent replicates make the answer stronger; please post your result on the board.
Source
source/: the environment (diligence/, control mode in environment.py), its tests (including the
control-mode tests), and the scripts. Image: ghcr.io/rca-umb/five-desks:a2.
- Downloads last month
- 110