Evaluate

View as Markdown

For the underlying concepts, read Evaluation and Environments. For server-level architecture, read Architecture.

NeMo Gym evaluation is environment-native: the same dataset, resources server, verifier, and rollout machinery supports model comparison, harness comparison, benchmark scoring, ablation studies, and training-data analysis.

Gym Evaluation Loop

  1. Discover — find an environment or benchmark that matches your target capability: gym list benchmarks, gym list environments, gym search <query>.
  2. Configure — choose a model and agent harness; pre-flight check your config with gym env validate.
  3. Run — collect rollouts with fixed sampling settings and an explicit repeat count: gym eval run. Resume an interrupted run with --resume.
  4. Scale — shard across jobs and merge outputs with gym eval aggregate.
  5. Profile — compute pass@1, pass@k, and per-task variance: gym eval profile.
  6. Diagnose — inspect failures, sometimes-pass tasks, and missing rows with BLADE.
  7. Decide — use results to determine whether the next action is model training, harness work, verifier repair, prompt change, or skill improvement.

The most important rule: vary one thing at a time. Changing the model, harness, prompt, verifier, and dataset together can still produce a useful release-gate score — but it cannot explain what caused the difference.

Models

NeMo Gym accesses models through the model server, which keeps provider-specific details behind the Responses API boundary — the same evaluation can compare hosted models, self-hosted vLLM instances, local checkpoints, or pre- and post-training snapshots.

Agent Harnesses

An agent harness is the orchestration layer that turns a model into an agent — it manages conversation state, routes tool calls, and decides when a task is complete. Harness changes often affect metrics as much as model changes do.

Benchmarks

A benchmark is a fixed evaluation configuration built on top of an environment’s resources server — same verifier, frozen dataset split, documented repeat count. All environments can be used for training; a benchmark is the evaluation-configured overlay.

Next Steps