Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Rulebook Desk

An OpenEnv environment for RL with verifiable rewards, built for the OpenEnv Arena. Container image: ghcr.io/skyforce/rulebook-desk:v1.

Each episode drops a fresh, procedurally generated workspace into /app: data files plus a short rulebook. The agent works through a shell using the Terminal-Bench 2 (tbench2_env) action schema (exec, write_file, evaluate), writes a machine-checkable answer, and calls evaluate once. The reward is partial credit in [0, 1].

What it teaches

Read an instruction and a rulebook, inspect files with the shell, script the computation instead of eyeballing it, apply stacked rules in the stated order (unit conversions, duplicates, caps, sliding windows, tie-breaks, rounding conventions), and hand in an answer in exactly the requested format. These are the habits that terminal tasks in every domain reward.

Tasks

tasks.jsonl lists the 32 task ids (16 families Γ— 2 levels, 4 per arena domain). Every reset generates a new instance, so the model cannot memorise answers; pass seed to reproduce one. samples/<task_id>/ holds the seed-0 workspace and instruction for each task.

Domain Families
Software engineering sw-repair (fix seeded bugs, hidden tests), sw-logs (access-log analytics)
Industrial / physical systems ind-oee (OEE from shift logs), ind-sensors (alarm events in mixed-unit sensor data)
Natural science sci-decay (half-life fit with dead-time correction), sci-weather (station climatology with QC)
Office / white-collar office-expenses (expense policy), office-schedule (cross-time-zone meeting slots)
Finance / economics fin-loan (amortisation with cent rounding), fin-fifo (FIFO gains, fees, splits)
Math / formal reasoning math-count (lattice paths, bounded compositions), math-modular (CRT, inverses, inclusion-exclusion)
Cybersecurity sec-authlog (SSH brute-force triage), sec-config-audit (Terraform + Kubernetes policy audit)
Media / content production media-subtitles (SRT repair), media-edl (EDL timecode maths)

Level 2 of each family stacks more rules (overlaps, unit mixes, out-of-order and superseded rows, caps, multi-document YAML, stock splits, non-coprime moduli, ...).

Reward

  • Answer files: share of checked fields correct; dict-valued fields count each key; lists are scored by F1; numeric tolerances are stated in the instruction or rulebook.
  • sw-repair: (share of defect-revealing hidden cases fixed) Γ— (share of regression cases still passing).
  • media-subtitles: share of expected cues reproduced exactly, over max(expected, produced).

Partial credit is deliberate: GRPO needs reward variance inside a group of rollouts, and all-or-nothing tasks give many zero-variance groups.

Integrity and evidence (evidence/)

  • The answer key and hidden tests stay in server memory. Commands run as an unprivileged agent user; the server code is root-only. Writes outside the workspace are refused.
  • controls_10seeds.txt: for 32 tasks Γ— 10 seeds, no answer β†’ 0, reference β†’ 1, one broken field β†’ strictly between 0 and 1, all fields wrong β†’ < 0.5, malformed JSON β†’ 0, untouched buggy module β†’ 0.
  • independent_solvers_25seeds.txt: solvers written separately from the generators, reading only the workspace and the instruction, score 1.0 on 30 tasks Γ— 25 seeds, so the files plus rulebook determine the key.
  • in_sandbox_solves.txt: the same solvers, uploaded with write_file and run with exec as the agent user inside the container, score 1.0 on fresh random instances.
  • validate_local.txt, replay_ws_local.txt: openenv validate --url passes 6/6; 97 WebSocket episodes over all tasks (example actions, finish action, reference answers, isolation checks, step limit) pass.
  • generation_sweep_500seeds.txt: 16,000 generated instances, no crash.

These are implementation controls, not results from a trained model.

Source

source/rulebook_desk/ is the environment as built into the image (OpenEnv revision 86a180ede21e044f7929b9a7783ad83aa67d83a3). The action/observation models copy OpenEnv's BSD-licensed tbench2_env field layout; the rest is original. Dataset: iskyforce/rulebook-desk.

Downloads last month
80