Glossary

Terms used in RL-environment and task-data procurement, each with a note on how it appears in a DataMining delivery.

RLVR (Reinforcement Learning with Verifiable Rewards)

RL where the reward comes from a programmatic checker rather than a learned judge. Our coding and agent environments are RLVR-compatible: the grader is deterministic code, so the reward channel cannot be influenced by the model.

F2P / P2P tests

Fail-to-Pass tests must fail before the correct patch and pass after; Pass-to-Pass tests must keep passing so fixes cannot break neighbors. We deliver both sets as private hidden tests, versioned per task.

pass@k

Probability of at least one correct solution in k attempts. We report per-task solve counts under a stated budget (for example, GPT-5.6 Sol xhigh, 4 valid attempts) instead of a bare pass@k, so the setup is checkable.

Reward hacking

The agent games the verifier — deleting tests, hardcoding outputs — instead of solving the task. Countermeasures in our environments: read-only test mounts, checksums, allow-listed networking, and deterministic grading. See our agent-environment checklist on GitHub.

Verifier

The program that decides whether a task attempt passes. Ours are deterministic, run locally, and ship with the task; a verifier that cannot be rerun by the buyer does not ship.

Trajectory

The full act–observe–revise log of one agent attempt. Deliveries include success and failure trajectories with honest labels — failures carry the gradient signal in RL.

Golden patch

The reference solution diff, stored separately from the model-visible package and never published with the task.

Contamination

A task entering model pretraining or post-training data, inflating scores. Countermeasures: original authoring, multi-model probes across families, item versioning with retirement.

Questions about a term in a delivery context: contact@datamining.cloud · RL environment delivery overview