Evaluate Environments

Evaluate Tasksets

Goal

Set up an evaluation for a taskset in the correct way to reproduce results from others or evaluate a model and harness combination on a given taskset.

Canonical path

Use the eval entrypoint

uv run eval <MY_ENV>

Core workflow

1. Resolve and validate config without model calls:

uv run eval <MY_ENV> --dry-run

2. Run model-free gold validation when the taskset implements validate:

uv run validate <MY_ENV> --runtime.type subprocess

3. Do a small run to see whether it works correctly:

uv run eval <MY_ENV> -m deepseek/deepseek-v4-flash -n 3 -r 1

4. Inspect successful, zero-reward, and errored traces.
5. Scale only after task loading, harness capability, runtime lifecycle, and scoring are correct.

When the user requests a full run, do not restrict the number of tasks. Ask for the appropriate harness to use (if not specified)

IDs and plugin resolution

A plugin id names an installed package (e.g. my-taskset); verifiers imports it and never installs anything itself.

The leading ID is shorthand for --env.taskset.id. A harness belongs to an agent — --env.agent.harness.* on the single-agent env, --env.<agent>.harness.* on a multi-agent one (there is no run-level --harness.*):

uv run eval my-task-v1 --env.agent.harness.id codex --env.agent.runtime.type prime

The env — the control flow between agents — owns the whole [env] block. Empty --env.id keeps the taskset's own story (its exported Env subclass, else the single-agent env); --env.id pairs a reusable env with any taskset, its knobs typed under --env.*:

uv run eval my-task-v1 --env.id best-of-n --env.n 8      # pass@k / rejection sampling
uv run eval my-task-v1 --env.id agentic-judge \
  --env.judge.runtime.type docker                           # a judge agent verifies each attempt in a sandbox

Disabling tools

Almost every harness comes with a disabled_tools list, which can be used to disable one or multiple tools:

[env.agent.harness]
disabled_tools = ["shell_tool"]

The names of these tools are set by the respective harness. Research the relevant first party documentation for the given harness for the relevant name(s). Some harnesses do not offer support to disable tools.

Config discovery

The CLI help is generated from the current config classes. Include the taskset and env ids you plan to use before --help so their concrete config fields are loaded:

uv run eval my-task-v1 \
  --env.id best-of-n \
  --help

For implementation details and defaults, start at verifiers/v1/configs/cli/eval.py and follow its fields into verifiers/v1/configs/. Client configs live in verifiers/v1/configs/client.py, sampling in verifiers/v1/types.py, and runtime- and harness-specific configs next to their implementations in verifiers/v1/runtimes/ and verifiers/v1/harnesses/. Custom taskset and env config fields live next to those implementations.

Typed taskset overrides

Taskset settings:

uv run eval my-task-v1 --env.taskset.split test --env.taskset.difficulty hard

Harness and runtime settings:

uv run eval my-task-v1 \
  --env.agent.harness.id rlm \
  --env.agent.runtime.type docker \
  --env.agent.runtime.cpu 4 \
  --env.agent.runtime.memory 8

Sampling:

uv run eval my-task-v1 \
  --sampling.temperature 0.7 \
  --sampling.top-p 0.95 \
  --sampling.max-tokens 2048 \
  --sampling.reasoning-effort medium

Always research the correct sampling parameters first. This is one of the most important settings, so make sure to find the correct values. For open models, you can find them on Hugging Face in the README and/or in the generation config.

Your parameter selection or settings should leave room for full runs, and you should not restrict things like tokens or number of turns unless specified by the user.

Leave optional settings unset unless the user asks for them. Always confirm the harness, runtime, and sampling parameters before running an evaluation.

Reproducible TOML

You can also use a TOML:

model = "openai/gpt-5-mini"

[env.taskset]
id = "my-task-v1"
split = "test"

[env.agent]
runtime = { type = "subprocess" }

[env.agent.harness]
id = "bash"

[sampling]
temperature = 0.7

uv run eval @ configs/my-eval.toml

Retries

Whole-rollout retry is opt-in. That means if something fails in the rollout, the whole rollout is retried. This is very useful for large-scale runs. You can also restrict certain errors from the retries:

uv run eval my-task-v1 \
  --env.agent.retries.max-retries 2 \
  --env.agent.retries.include SandboxError ProviderError \
  --env.agent.retries.exclude TaskError

Output and resume

A run writes to output_dir / run.dir (-o sets output_dir, default outputs; run.dir defaults to the auto-generated run name):

outputs/<env>--<model>--<harness>--<short-id>/
├── configs/eval.json
├── logs/eval.log
└── traces.jsonl

configs/eval.json is the run's resolved config, re-runnable via @. traces.jsonl is one episode per line — the episode's traces plus their shared standing — appended after each episode finishes, so an episode is durable whole or not at all (a torn last line is the whole episode redone on resume).

Resume in place by re-running the run's own saved config with --resume (it re-runs only the missing/errored rollouts; any config drift from the saved run is refused):

uv run eval @ <run-dir>/configs/eval.json --resume

To overwrite a run dir and start fresh instead, use --clean.

Trace inspection

For each representative sample inspect:

• task and prompt fields;
• branches, assistant messages, tool messages, and stop condition;
• named rewards, aggregate reward, and metrics;
• persisted info artifacts;
• error/errors and boundary type;
• per-call calls records (model, sampling, finish reason, usage, timing, error) linked to the graph;
• usage and stage timing;
• token/mask/logprob fields when using the training client.

Classify outcomes:

1. Valid completion and correct reward.
2. Valid completion with low reward (model/task outcome).
3. Truncated completion (budget outcome).
4. Captured rollout error (provider, harness, tool, user, runtime, task, or interception).

Do not average these categories together without reporting failure rate.

Metrics interpretation

• Binary rewards support solve rate and pass@k-style analysis.
• Continuous rewards need distributions, quantiles, and per-task/group comparisons.
• Group rewards must be interpreted with their comparison rule and group size.
• Always inspect samples before attributing a delta to model quality.
• Keep taskset, harness, runtime, sampling, and selected task indices fixed across variants.
• Do not overinterpret a tiny smoke run.