Envsmithy  ·  RL environments & training data

Environments,
forged at scale.

Verifiable RL environments and the training data that goes with them. Executable tasks, graders tested against wrong answers, trajectories with provenance.

gridworld demo
episodes 0 success mean return env gw-01

Environment suites

Three things, made to be trained on

Most available environments are either toy-sized or can't be checked. We build the third kind: realistic, verifiable, and released as suites.

Environments

Verifiable tasks

Each task carries an initial state, a tool interface, a success specification, validated solution traces, and an executable grader tested on valid, invalid, and near-miss outcomes.

Data

High-quality trajectories

Timestamped tool-use trajectories and pairwise preference records. Each sample carries annotator qualifications, task and environment versions, and review decisions.

Scale

Suites, not one-offs

Parameterised task variants generated from a domain spec, filtered by automated and expert checks, with difficulty estimated from success rates across named model-and-scaffold baselines.

Domains

Supported domains

Long-horizon work where success can be checked against state, not judged from the transcript.

CODING

Coding

Repository tasks graded on held-out functional tests, regression tests, and forbidden side effects. Patch similarity is not part of the reward.

COMPUTER USE

Computer & browser use

Multi-step tasks in real applications, verified on the resulting application state rather than the action sequence.

DATA ENGINEERING

Data engineering

Pipelines, schemas, and messy inputs. Outputs checked against ground truth; runs must be reproducible.

ENGINEERING DESIGN

Engineering design

Design tasks graded against explicit constraints with simulation or optimisation, plus expert review for the assumptions a solver doesn't capture.

Research

What every task is held to

Environment design is a research problem. Three properties are non-negotiable.

# fields every task record carries
task          id · domain · spec version
state         seeded initial state · deterministic reset
tools         shell · python · domain simulator
grader        deterministic state checks + rubric dimensions
validation    reference traces pass · near-miss + exploit probes fail
baseline      pass@1 per named model + scaffold, N trials
provenance    author · reviewers · review decisions

Record schema · values are placeholders, not measurements

Verifiable

Reward is decided by state checks or a solver. Rubric dimensions only where judgement is unavoidable, with inter-rater agreement measured and reported.

Non-trivial

Tasks solvable without interacting with the environment are rejected. Prompt-only and tool-enabled baselines are both reported.

Hack-resistant

Graders are tested against near-miss and exploit attempts before release, so the shortest path to reward is the intended one.