Attoloop/club-software

Club Software

15 software-engineering tasks for the OpenEnv Arena, toned down from three Attoloop environments so that a 27B model succeeds some of the time. Each task is a real working environment with its original grader; the easier versions change only the starting state, never the grading.

Source	Tasks	What the model must do	How it is toned down
dialect-recon	7	Make an interpreter match a hidden dialect of a small language, using 40 examples and a metered probe	A working base-language interpreter is given; its dialect settings are a dictionary with the allowed values listed. Rung r1 has all changes but one set, r2 the ones the examples reveal, r3 none
table-press	4	Encode HTTP/3 header lists in QPACK so two independent decoders reproduce them, in as few bytes as possible	Traces cut to 5 header lists; in e5 a valid literal-only encoder is given and the job is to make its output smaller
distribution-robust-sql-rewrite	4	Rewrite a PostgreSQL query so it returns exactly the same rows at a lower planner cost on hidden databases	The easiest band (predicate pushdown)

Tasks are listed in training order: the most-helped rungs of every source first, the least-helped last.

Interface

One episode is one task in a fresh container. The model replies with exactly one JSON object per turn:

• {"op": "run", "command": "..."} runs one shell command in the working directory.
• {"op": "write", "path": "...", "content": ["line", "line"]} writes a file, content given as a list of lines.
• {"op": "edit", "path": "...", "old": "...", "new": "..."} replaces one exact snippet that occurs once.
• {"op": "submit"} ends the episode; the task's own verifier grades the work.

The model's commands run as an unprivileged user; the verifier's files stay root-only until grading.

Reward

Each verifier's raw score is mapped into [0, 1] with a hyperbolic calibration between two anchors per task (tasks.jsonl lists them): the low anchor scores 0, the high anchor 0.4, and the curve keeps rising toward 1.0 above it.

• dialect-recon: 0 to 300 hidden programs matched (a submission failing the visible-examples gate scores 0).
• table-press: the low anchor is the gate floor, so any valid encoding earns reward; the high anchor is 0.32 times the reference literal encoding's size, the median ratio a frontier agent reached on the full traces.
• SQL rewrite: the grader's own quality score, 0 to 1.

Limitations

• The original tasks are sealed offline; the arena sandbox has outbound internet, which an unprivileged process cannot be denied. The adapter points the usual proxy variables at a closed port.
• The rungs' difficulty was set from the tasks' own data, not measured on the target model.

Licences

The tasks and verifiers are released under the Apache License 2.0 (LICENSE). Third-party material keeps its own licence; notices are in licenses/<environment>/ and inside each image.