nemo-evaluator

What it is and what it does

NeMo Evaluator is an LLM evaluation framework that runs standardized benchmarks (mmlu, mmlu_pro, gpqa, gsm8k, math500, mgsm, drop, triviaqa, humaneval, simpleqa, healthbench, pinchbench, xstest, terminal-bench-hard, terminal-bench-v1, nmp_harbor) against language models and produces scored results. It sits between your test harness and an LLM API, intercepting requests to cache responses, limit turns, modify payloads, and inject system messages—all without external proxy infrastructure. You define benchmarks in YAML or Python, choose a solver type (simple, harbor, tool_calling, gym_delegation, openclaw, container), and run evaluations locally or on SLURM clusters with Docker sandboxes for code execution.

The framework handles scoring (multichoice_regex, numeric_match, answer_line, fuzzy_match, code_sandbox, needs_judge), result aggregation, and export to experiment trackers or Inspect AI format. It ships with CLI commands to run benchmarks, merge sharded results, generate reports, compare runs, and enforce quality gates. Optional extras add symbolic math, statistical analysis, Harbor agent support, and Inspect export.

Use it for:

• Run mmlu or gsm8k against a model API to measure instruction-following and math reasoning performance.
• Intercept and cache LLM API calls during iterative benchmark development to reduce API costs.
• Execute code-based benchmarks (humaneval) in isolated Docker containers with automatic sandbox management.
• Compare baseline vs. candidate model runs with statistical deltas and pass/fail flips using nel compare.
• Enforce multi-benchmark quality gates with explicit pass/fail policies before production deployment.
• Export evaluation results to experiment trackers for team collaboration and result tracking.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

NeMo Evaluator runs standardized benchmarks against language models through a configurable framework with built-in scoring, proxy interception, and multi-format result reporting.

Yes, if you need to benchmark language models against standard datasets. Low install friction, active maintenance, no security vulnerabilities, and permissive license. Requires Python 3.12 or 3.13, an LLM endpoint, and familiarity with YAML configuration. Best suited for teams evaluating model performance systematically; less relevant for single ad-hoc inference tasks.

Install

nemo-evaluator on PyPI

pip install nemo-evaluator

uv add nemo-evaluator

poetry add nemo-evaluator

Installing nemo-evaluator

Before you install

Low friction: pure Python wheel with no compiled dependencies. Active maintenance (last commit 2026-08-13) and recent release cycle. Requires Python 3.12 or 3.13; optional extras add scipy, sympy, or agent integrations as needed.

License in practice

Apache 2.0 permissive license allows commercial and private use with minimal restrictions; attribution required but no copyleft obligations.

Quickstart

pip install nemo-evaluator
export NVIDIA_API_KEY="your-key"
nel eval run --bench mmlu --model-url https://integrate.api.nvidia.com/v1 --model-id nvidia/nemotron-3-super-120b-a12b --api-key $NVIDIA_API_KEY

Requires Python 3.12 or 3.13; most benchmarks need an external LLM API key or local model endpoint.

Verify before relying

• Whether all built-in benchmarks work out-of-the-box or require additional setup per benchmark type.
• Performance characteristics and typical runtime for standard benchmarks at scale.
• Compatibility of external harness integrations with current versions of lm-eval, skills, vlmevalkit, gym, harbor, and container systems.

Package facts

Evidence: nemo_evaluator-0.3.0-py3-none-any.whl

Keywords: LLM, NLP, NVIDIA, benchmarks, deep learning, evaluations, gpu, machine learning, pytorch

Tags

More Artificial Intelligence packages

Further reading