nemo-evaluator（尼莫评估器）

NeMo Evaluator SDK - Enterprise LLM Benchmarking

Quick Start

NeMo Evaluator SDK evaluates LLMs across 100+ benchmarks from 18+ harnesses using containerized, reproducible evaluation with multi-backend execution (local Docker, Slurm HPC, Lepton cloud).

Installation:

pip install nemo-evaluator-launcher

Basic evaluation:

export NGC_API_KEY=nvapi-your-key-here

cat > config.yaml << 'EOF'
defaults:
  - execution: local
  - deployment: none
  - _self_

execution:
  output_dir: ./results

target:
  api_endpoint:
    model_id: meta/llama-3.1-8b-instruct
    url: https://integrate.api.nvidia.com/v1/chat/completions
    api_key_name: NGC_API_KEY

evaluation:
  tasks:
    - name: ifeval
EOF

nemo-evaluator-launcher run --config-dir . --config-name config

Common Workflows

Workflow 1: Standard Model Evaluation

Checklist:

- [ ] Configure API endpoint (NVIDIA Build or self-hosted)
- [ ] Select benchmarks (MMLU, GSM8K, IFEval, HumanEval)
- [ ] Run evaluation
- [ ] Check results

Step 1: Configure endpoint

For NVIDIA Build:

target:
  api_endpoint:
    model_id: meta/llama-3.1-8b-instruct
    url: https://integrate.api.nvidia.com/v1/chat/completions
    api_key_name: NGC_API_KEY

For self-hosted (vLLM, TRT-LLM):

target:
  api_endpoint:
    model_id: my-model
    url: http://localhost:8000/v1/chat/completions
    api_key_name: ""

Step 2: Select benchmarks

evaluation:
  tasks:
    - name: ifeval           # Instruction following
    - name: gpqa_diamond     # Graduate-level QA
      env_vars:
        HF_TOKEN: HF_TOKEN
    - name: gsm8k_cot_instruct  # Math reasoning
    - name: humaneval        # Code generation

Step 3: Run and check results

nemo-evaluator-launcher run --config-dir . --config-name config
nemo-evaluator-launcher status <invocation_id>
cat results/<invocation_id>/<task>/artifacts/results.yml

Workflow 2: Slurm HPC Evaluation

defaults:
  - execution: slurm
  - deployment: vllm
  - _self_

execution:
  hostname: cluster.example.com
  account: my_slurm_account
  partition: gpu
  output_dir: /shared/results
  walltime: "04:00:00"
  nodes: 1
  gpus_per_node: 8

deployment:
  checkpoint_path: /shared/models/llama-3.1-8b
  tensor_parallel_size: 2
  data_parallel_size: 4

Workflow 3: Model Comparison

# Same config, different models
nemo-evaluator-launcher run --config-dir . --config-name base_eval \
  -o target.api_endpoint.model_id=meta/llama-3.1-8b-instruct

nemo-evaluator-launcher run --config-dir . --config-name base_eval \
  -o target.api_endpoint.model_id=mistralai/mistral-7b-instruct-v0.3

# Export results
nemo-evaluator-launcher export <id> --dest mlflow
nemo-evaluator-launcher export <id> --dest wandb

Supported Harnesses

CLI Reference

When to Use vs Alternatives

Use NeMo Evaluator when:

• Need 100+ benchmarks from 18+ harnesses
• Running on Slurm HPC clusters
• Requiring reproducible containerized evaluation
• Evaluating against OpenAI-compatible APIs

Use alternatives instead:

• lm-evaluation-harness: Simpler local evaluation
• bigcode-evaluation-harness: Code-only benchmarks
• HELM: Broader evaluation (fairness, efficiency)

Common Issues

Container pull fails: Configure NGC credentials

docker login nvcr.io -u '$oauthtoken' -p $NGC_API_KEY

Task requires env var: Add to task config

tasks:
  - name: gpqa_diamond
    env_vars:
      HF_TOKEN: HF_TOKEN

Increase parallelism:

-o +evaluation.nemo_evaluator_config.config.params.parallelism=8
-o +evaluation.nemo_evaluator_config.config.params.limit_samples=100

Requirements

• Python 3.10-3.13
• Docker (for local execution)
• NGC API Key (for NVIDIA Build)
• HF_TOKEN (for some benchmarks)