NeMo Evaluator SDK - Enterprise LLM Benchmarking

# 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**:
```bash
pip install nemo-evaluator-launcher
```

**Set API key and run evaluation**:
```bash
export NGC_API_KEY=nvapi-your-key-here

# Create minimal config
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

# Run evaluation
nemo-evaluator-launcher run --config-dir . --config-name config
```

**View available tasks**:
```bash
nemo-evaluator-launcher ls tasks
```

## Common Workflows

### Workflow 1: Evaluate Model on Standard Benchmarks

Run core academic benchmarks (MMLU, GSM8K, IFEval) on any OpenAI-compatible endpoint.

**Checklist**:
```
Standard Evaluation:
- [ ] Step 1: Configure API endpoint
- [ ] Step 2: Select benchmarks
- [ ] Step 3: Run evaluation
- [ ] Step 4: Check results
```

**Step 1: Configure API endpoint**

```yaml
# config.yaml
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
```

For self-hosted endpoints (vLLM, TRT-LLM):
```yaml
target:
  api_endpoint:
    model_id: my-model
    url: http://localhost:8000/v1/chat/completions
    api_key_name: ""  # No key needed for local
```

**Step 2: Select benchmarks**

Add tasks to your config:
```yaml
evaluation:
  tasks:
    - name: ifeval           # Instruction following
    - name: gpqa_diamond     # Graduate-level QA
      env_vars:
        HF_TOKEN: HF_TOKEN   # Some tasks need HF token
    - name: gsm8k_cot_instruct  # Math reasoning
    - name: humaneval        # Code generation
```

**Step 3: Run evaluation**

```bash
# Run with config file
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name config

# Override output directory
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name config \
  -o execution.output_dir=./my_results

# Limit samples for quick testing
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name config \
  -o +evaluation.nemo_evaluator_config.config.params.limit_samples=10
```

**Step 4: Check results**

```bash
# Check job status
nemo-evaluator-launcher status <invocation_id>

# List all runs
nemo-evaluator-launcher ls runs

# View results
cat results/<invocation_id>/<task>/artifacts/results.yml
```

### Workflow 2: Run Evaluation on Slurm HPC Cluster

Execute large-scale evaluation on HPC infrastructure.

**Checklist**:
```
Slurm Evaluation:
- [ ] Step 1: Configure Slurm settings
- [ ] Step 2: Set up model deployment
- [ ] Step 3: Launch evaluation
- [ ] Step 4: Monitor job status
```

**Step 1: Configure Slurm settings**

```yaml
# slurm_config.yaml
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
```

**Step 2: Set up model deployment**

```yaml
deployment:
  checkpoint_path: /shared/models/llama-3.1-8b
  tensor_parallel_size: 2
  data_parallel_size: 4
  max_model_len: 4096

target:
  api_endpoint:
    model_id: llama-3.1-8b
    # URL auto-generated by deployment
```

**Step 3: Launch evaluation**

```bash
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name slurm_config
```

**Step 4: Monitor job status**

```bash
# Check status (queries sacct)
nemo-evaluator-launcher status <invocation_id>

# View detailed info
nemo-evaluator-launcher info <invocation_id>

# Kill if needed
nemo-evaluator-launcher kill <invocation_id>
```

### Workflow 3: Compare Multiple Models

Benchmark multiple models on the same tasks for comparison.

**Checklist**:
```
Model Comparison:
- [ ] Step 1: Create base config
- [ ] Step 2: Run evaluations with overrides
- [ ] Step 3: Export and compare results
```

**Step 1: Create base config**

```yaml
# base_eval.yaml
defaults:
  - execution: local
  - deployment: none
  - _self_

execution:
  output_dir: ./comparison_results

evaluation:
  nemo_evaluator_config:
    config:
      params:
        temperature: 0.01
        parallelism: 4
  tasks:
    - name: mmlu_pro
    - name: gsm8k_cot_instruct
    - name: ifeval
```

**Step 2: Run evaluations with model overrides**

```bash
# Evaluate Llama 3.1 8B
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name base_eval \
  -o target.api_endpoint.model_id=meta/llama-3.1-8b-instruct \
  -o target.api_endpoint.url=https://integrate.api.nvidia.com/v1/chat/completions

# Evaluate Mistral 7B
nemo-evaluator-launcher run \
  --config-dir . \
  --config-name base_eval \
  -o target.api_endpoint.model_id=mistralai/mistral-7b-instruct-v0.3 \
  -o target.api_endpoint.url=https://integrate.api.nvidia.com/v1/chat/completions
```

**Step 3: Export and compare**

```bash
# Export to MLflow
nemo-evaluator-launcher export <invocation_id_1> --dest mlflow
nemo-evaluator-launcher export <invocation_id_2> --dest mlflow

# Export to local JSON
nemo-evaluator-launcher export <invocation_id> --dest local --format json

# Export to Weights & Biases
nemo-evaluator-launcher export <invocation_id> --dest wandb
```

### Workflow 4: Safety and Vision-Language Evaluation

Evaluate models on safety benchmarks and VLM tasks.

**Checklist**:
```
Safety/VLM Evaluation:
- [ ] Step 1: Configure safety tasks
- [ ] Step 2: Set up VLM tasks (if applicable)
- [ ] Step 3: Run evaluation
```

**Step 1: Configure safety tasks**

```yaml
evaluation:
  tasks:
    - name: aegis              # Safety harness
    - name: wildguard          # Safety classification
    - name: garak              # Security probing
```

**Step 2: Configure VLM tasks**

```yaml
# For vision-language models
target:
  api_endpoint:
    type: vlm  # Vision-language endpoint
    model_id: nvidia/llama-3.2-90b-vision-instruct
    url: https://integrate.api.nvidia.com/v1/chat/completions

evaluation:
  tasks:
    - name: ocrbench           # OCR evaluation
    - name: chartqa            # Chart understanding
    - name: mmmu               # Multimodal understanding
```

## When to Use vs Alternatives

**Use NeMo Evaluator when:**
- Need **100+ benchmarks** from 18+ harnesses in one platform
- Running evaluations on **Slurm HPC clusters** or cloud
- Requiring **reproducible** containerized evaluation
- Evaluating against **OpenAI-compatible APIs** (vLLM, TRT-LLM, NIMs)
- Need **enterprise-grade** evaluation with result export (MLflow, W&B)

**Use alternatives instead:**
- **lm-evaluation-harness**: Simpler setup for quick local evaluation
- **bigcode-evaluation-harness**: Focused only on code benchmarks
- **HELM**: Stanford's broader evaluation (fairness, efficiency)
- **Custom scripts**: Highly specialized domain evaluation

## Supported Harnesses and Tasks

| Harness | Task Count | Categories |
|---------|-----------|------------|
| `lm-evaluation-harness` | 60+ | MMLU, GSM8K, HellaSwag, ARC |
| `simple-evals` | 20+ | GPQA, MATH, AIME |
| `bigcode-evaluation-harness` | 25+ | HumanEval, MBPP, MultiPL-E |
| `safety-harness` | 3 | Aegis, WildGuard |
| `garak` | 1 | Security probing |
| `vlmevalkit` | 6+ | OCRBench, ChartQA, MMMU |
| `bfcl` | 6 | Function calling v2/v3 |
| `mtbench` | 2 | Multi-turn conversation |
| `livecodebench` | 10+ | Live coding evaluation |
| `helm` | 15 | Medical domain |
| `nemo-skills` | 8 | Math, science, agentic |

## Common Issues

**Issue: Container pull fails**

Ensure NGC credentials are configured:
```bash
docker login nvcr.io -u '$oauthtoken' -p $NGC_API_KEY
```

**Issue: Task requires environment variable**

Some tasks need HF_TOKEN or JUDGE_API_KEY:
```yaml
evaluation:
  tasks:
    - name: gpqa_diamond
      env_vars:
        HF_TOKEN: HF_TOKEN  # Maps env var name to env var
```

**Issue: Evaluation timeout**

Increase parallelism or reduce samples:
```bash
-o +evaluation.nemo_evaluator_config.config.params.parallelism=8
-o +evaluation.nemo_evaluator_config.config.params.limit_samples=100
```

**Issue: Slurm job not starting**

Check Slurm account and partition:
```yaml
execution:
  account: correct_account
  partition: gpu
  qos: normal  # May need specific QOS
```

**Issue: Different results than expected**

Verify configuration matches reported settings:
```yaml
evaluation:
  nemo_evaluator_config:
    config:
      params:
        temperature: 0.0  # Deterministic
        num_fewshot: 5    # Check paper's fewshot count
```

## CLI Reference

| Command | Description |
|---------|-------------|
| `run` | Execute evaluation with config |
| `status <id>` | Check job status |
| `info <id>` | View detailed job info |
| `ls tasks` | List available benchmarks |
| `ls runs` | List all invocations |
| `export <id>` | Export results (mlflow/wandb/local) |
| `kill <id>` | Terminate running job |

## Configuration Override Examples

```bash
# Override model endpoint
-o target.api_endpoint.model_id=my-model
-o target.api_endpoint.url=http://localhost:8000/v1/chat/completions

# Add evaluation parameters
-o +evaluation.nemo_evaluator_config.config.params.temperature=0.5
-o +evaluation.nemo_evaluator_config.config.params.parallelism=8
-o +evaluation.nemo_evaluator_config.config.params.limit_samples=50

# Change execution settings
-o execution.output_dir=/custom/path
-o execution.mode=parallel

# Dynamically set tasks
-o 'evaluation.tasks=[{name: ifeval}, {name: gsm8k}]'
```

## Python API Usage

For programmatic evaluation without the CLI:

```python
from nemo_evaluator.core.evaluate import evaluate
from nemo_evaluator.api.api_dataclasses import (
    EvaluationConfig,
    EvaluationTarget,
    ApiEndpoint,
    EndpointType,
    ConfigParams
)

# Configure evaluation
eval_config = EvaluationConfig(
    type="mmlu_pro",
    output_dir="./results",
    params=ConfigParams(
        limit_samples=10,
        temperature=0.0,
        max_new_tokens=1024,
        parallelism=4
    )
)

# Configure target endpoint
target_config = EvaluationTarget(
    api_endpoint=ApiEndpoint(
        model_id="meta/llama-3.1-8b-instruct",
        url="https://integrate.api.nvidia.com/v1/chat/completions",
        type=EndpointType.CHAT,
        api_key="nvapi-your-key-here"
    )
)

# Run evaluation
result = evaluate(eval_cfg=eval_config, target_cfg=target_config)
```

## Advanced Topics

**Multi-backend execution**: See [references/execution-backends.md](references/execution-backends.md)
**Configuration deep-dive**: See [references/configuration.md](references/configuration.md)
**Adapter and interceptor system**: See [references/adapter-system.md](references/adapter-system.md)
**Custom benchmark integration**: See [references/custom-benchmarks.md](references/custom-benchmarks.md)

## Requirements

- **Python**: 3.10-3.13
- **Docker**: Required for local execution
- **NGC API Key**: For pulling containers and using NVIDIA Build
- **HF_TOKEN**: Required for some benchmarks (GPQA, MMLU)

## Resources

- **GitHub**: https://github.com/NVIDIA-NeMo/Evaluator
- **NGC Containers**: nvcr.io/nvidia/eval-factory/
- **NVIDIA Build**: https://build.nvidia.com (free hosted models)
- **Documentation**: https://github.com/NVIDIA-NeMo/Evaluator/tree/main/docs