EvalHub SDK

Framework Adapter SDK for EvalHub Integration

The EvalHub SDK provides a standardized way to create framework adapters that can be consumed by EvalHub, enabling a "Bring Your Own Framework" (BYOF) approach for evaluation frameworks.

Overview

The SDK creates a common API layer that allows EvalHub to communicate with ANY evaluation framework. Users only need to write minimal "glue" code to connect their framework to the standardized interface.

EvalHub → (Standard API) → Your Framework Adapter → Your Evaluation Framework

Architecture

The adapter SDK uses a job runner architecture:

graph TB
    subgraph pod["Kubernetes Job Pod"]
        subgraph adapter["Adapter Container"]
            A1["1. Read JobSpec<br/>from ConfigMap"]
            A2["2. run_benchmark_job()"]
            A3["3. Report status<br/>via callbacks"]
            A4["4. Create OCI artifacts<br/>via callbacks"]
            A5["5. Report results<br/>via callbacks"]
            A6["6. Exit"]
        end

        subgraph sidecar["Sidecar Container"]
            S1["ConfigMap mounted<br/>/meta/job.json"]
            S2["Forward status to<br/>EvalHub service (HTTP)"]
            S3["Authenticated push of<br/>OCI artifacts<br/>to OCI Registry"]
            S4["Forward results to<br/>EvalHub service (HTTP)"]
        end

        A1 -.-> S1
        A3 --> S2
        A4 --> S3
        A5 --> S4
    end

    S2 --> EvalHub["EvalHub Service"]
    S3 --> Registry["OCI Registry"]
    S4 --> EvalHub

    style pod fill:#f0f0f0,stroke:#333,stroke-width:2px
    style adapter fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    style sidecar fill:#fff3e0,stroke:#f57c00,stroke-width:2px

Package Organization

The SDK is organized into distinct, focused packages:

Core (evalhub.models) - Shared data models

• Request/response models for API communication
• Common data structures for evaluations and benchmarks

Adapter SDK (evalhub.adapter) - Framework adapter components

• FrameworkAdapter base class with run_benchmark_job() method
• Job specification models (JobSpec, JobResults)
• Callback interface for status updates and OCI artifacts
• Example implementations

Client SDK (evalhub.client) - REST API client for EvalHub service

• HTTP client for submitting evaluations to EvalHub
• Job lifecycle: submit, status, cancel, wait, and log fetch/watch
• Resource navigation (providers, benchmarks, collections)
• See Getting Started with the CLI

Key Components

1. JobSpec - Job configuration loaded from ConfigMap at pod startup
2. FrameworkAdapter - Base class that implements run_benchmark_job() method
3. JobCallbacks - Interface for reporting status and persisting artifacts
4. JobResults - Evaluation results returned when job completes
5. EvalCardMetadata - Standardized evaluation disclosure (Dhar et al., arXiv:2511.21695): modalities, languages, capability and safety evaluations
6. EnvironmentCardMetadata - Operational context of an evaluation run: hardware, software, Kubernetes, model identity, and run provenance
7. additional_info - Supplementary key-value pairs for evaluation information beyond metrics (e.g. dataset provenance, zero-shot/alt-prompting scores)
8. Sidecar - Container that handles service communication (provided by platform)

Breaking Changes

MCP command (> 0.4.1)

The evalhub mcp command no longer starts a stdio MCP server directly. MCP hosts (Claude, etc.) that previously used evalhub mcp must switch to evalhub mcp run. The old evalhub mcp interface works only on versions <= 0.4.1.

Quick Start

1. Installation

# Install from PyPI (when available)
pip install eval-hub-sdk

# Install from source
git clone https://github.com/eval-hub/eval-hub-sdk.git
cd eval-hub-sdk
pip install -e .[dev]

Optional extras: core, adapter, client, cli, mcp, dev, server, all. The server extra bundles the eval-hub-server binary for local/embedded use and is excluded from all due to its size (~20 MB).

2. Create Your Adapter

Create a new Python file for your adapter:

# my_framework_adapter.py
from datetime import UTC, datetime
from pathlib import Path

from evalhub.adapter import (
    FrameworkAdapter,
    JobSpec,
    JobCallbacks,
    JobResults,
    JobStatus,
    JobPhase,
    JobStatusUpdate,
    EvaluationResult,
    MessageInfo,
    OCIArtifactSpec,
)

class MyFrameworkAdapter(FrameworkAdapter):
    def run_benchmark_job(
        self, config: JobSpec, callbacks: JobCallbacks
    ) -> JobResults:
        """Run a benchmark evaluation job."""

        # Report initialization
        callbacks.report_status(JobStatusUpdate(
            status=JobStatus.RUNNING,
            phase=JobPhase.INITIALIZING,
            progress=0.0,
            message=MessageInfo(
                message="Loading benchmark and model",
                message_code="initializing",
            ),
        ))

        # Load your evaluation framework and benchmark
        framework = load_your_framework()
        benchmark = framework.load_benchmark(config.benchmark_id)
        model = framework.load_model(config.model)

        # Report evaluation start
        callbacks.report_status(JobStatusUpdate(
            status=JobStatus.RUNNING,
            phase=JobPhase.RUNNING_EVALUATION,
            progress=0.3,
            message=MessageInfo(
                message=f"Evaluating on {config.num_examples} examples",
                message_code="running_evaluation",
            ),
        ))

        # Run evaluation (adapter-specific params come from parameters)
        results = framework.evaluate(
            benchmark=benchmark,
            model=model,
            num_examples=config.num_examples,
            num_few_shot=config.parameters.get("num_few_shot", 0)
        )

        # Save results to a directory and persist as OCI artifact
        results_dir = save_results(config.id, results)
        oci_artifact = None
        oci_exports = config.exports.oci if config.exports else None
        if oci_exports is not None:
            coords = oci_exports.coordinates.model_copy(deep=True)
            coords.annotations.update({
                "org.opencontainers.image.created": datetime.now(UTC).isoformat(),
                "io.github.eval-hub.benchmark": config.benchmark_id,
                "io.github.eval-hub.model": config.model.name,
                "io.github.eval-hub.job_id": config.id,
            })
            oci_artifact = callbacks.create_oci_artifact(OCIArtifactSpec(
                files_path=results_dir,
                coordinates=coords,
            ))

        # Return results
        return JobResults(
            id=config.id,
            benchmark_id=config.benchmark_id,
            benchmark_index=config.benchmark_index,
            model_name=config.model.name,
            results=[
                EvaluationResult(
                    metric_name="accuracy",
                    metric_value=results["accuracy"],
                    metric_type="float"
                )
            ],
            num_examples_evaluated=len(results),
            duration_seconds=results["duration"],
            oci_artifact=oci_artifact,
        )

3. OCI Artifact Persistence

The SDK exposes an OCI persistence API via callbacks.create_oci_artifact(...).

Using DefaultCallbacks

Use DefaultCallbacks for both production and development:

from evalhub.adapter import DefaultCallbacks

# Initialize adapter (loads settings and job spec internally)
adapter = MyFrameworkAdapter()

# Create callbacks from adapter (auto-configures sidecar, OCI proxy, etc.)
callbacks = DefaultCallbacks.from_adapter(adapter)

results = adapter.run_benchmark_job(adapter.job_spec, callbacks)

Key Points:

• Status updates: Sent to sidecar if sidecar_url is provided, otherwise logged locally. Both report_status and report_results events always include benchmark_index (and provider_id when set) so the service can associate events with the correct benchmark in multi-benchmark jobs.
• OCI artifacts: Created via SDK callbacks and pushed to the OCI registry through the sidecar-authenticated flow when mode is Kubernetes.

4. Containerise Your Adapter

Create a Dockerfile for your adapter:

FROM registry.access.redhat.com/ubi9/python-312

WORKDIR /app

# Install dependencies
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Copy adapter code
COPY my_framework_adapter.py .
COPY run_adapter.py .

# Run adapter
CMD ["python", "run_adapter.py"]

Create the entrypoint script:

# run_adapter.py
from my_framework_adapter import MyFrameworkAdapter
from evalhub.adapter import DefaultCallbacks

# Initialize adapter (loads settings and job spec internally)
adapter = MyFrameworkAdapter()

# Create callbacks from adapter (auto-configures sidecar, OCI proxy, etc.)
callbacks = DefaultCallbacks.from_adapter(adapter)

# Run adapter
results = adapter.run_benchmark_job(adapter.job_spec, callbacks)

# Report final results to service via sidecar
callbacks.report_results(results)

print(f"Job completed: {results.id}")

5. Deploy to Kubernetes

The eval-hub service will create Kubernetes Jobs for your adapter:

apiVersion: batch/v1
kind: Job
metadata:
  name: eval-job-123
spec:
  template:
    spec:
      containers:
      # Your adapter container
      - name: adapter
        image: myregistry/my-adapter:latest
        volumeMounts:
        - name: job-spec
          mountPath: /meta
      # Sidecar container (provided by platform)
      - name: sidecar
        image: evalhub/sidecar:latest
        env:
        - name: EVALHUB_SERVICE_URL
          value: "http://evalhub-service:8080"
      volumes:
      - name: job-spec
        configMap:
          name: job-123-spec

For a complete working example, see examples/simple_adapter/simple_adapter.py.

Package Organization Guide

The EvalHub SDK is organized into distinct packages based on your use case:

Which Package Should I Use?

Import Patterns

Framework Adapter Developer:

# Building your adapter
from evalhub.adapter import (
    FrameworkAdapter,
    JobSpec,
    JobCallbacks,
    JobResults,
    JobStatus,
    JobPhase,
    JobStatusUpdate,
    EvaluationResult,
    OCIArtifactSpec,
    # Card metadata (optional — auto-capture provides a baseline)
    CapabilityEvalEntry,
    EvalCardMetadata,
    EnvironmentCardMetadata,
)

EvalHub Service User:

# Interacting with EvalHub REST API
from evalhub import (
    JobLogOptions,
    SyncEvalHubClient,
    BenchmarkConfig,
    EvaluationExports,
    EvaluationExportsOCI,
    JobLogOptions,
    JobSubmissionRequest,
    ModelConfig,
    OCIConnectionConfig,
    OCICoordinates,
)

# Watch job logs while polling status until the job completes
with SyncEvalHubClient() as client:
    for update in client.jobs.watch_logs(
        "job-id",
        options=JobLogOptions(tail_lines=500),
        poll_interval=2.0,
    ):
        if update.logs:
            print(update.logs, end="")

Examples

Watch job logs

Stream workload logs while a job runs using the client API or the example script:

• Script: examples/watch_job_logs.py — runnable against a local or remote cluster; see examples/README.md
• Client API: client.jobs.get_logs() for a one-shot snapshot; client.jobs.watch_logs() to poll logs and status until the job reaches a terminal state (yields JobLogUpdate with incremental logs and current job)

Contributed Adapters

For real use-case adapter implementations, see the eval-hub-contrib repository which includes adapters for GuideLLM, LightEval, and MTEB.

Simple Adapter Example

The SDK includes a reference implementation showing all adapter patterns:

Example Adapter: examples/simple_adapter/simple_adapter.py

This example demonstrates:

• Loading JobSpec from mounted ConfigMap
• Validating configuration
• Loading benchmark data
• Running evaluation with progress reporting
• Persisting results as OCI artifacts
• Returning structured results

Using the Example

from evalhub.adapter.examples import ExampleAdapter
from evalhub.adapter import JobSpec

# Load job specification
job_spec = JobSpec(
    id="eval-123",
    provider_id="my-provider",
    benchmark_id="mmlu",
    benchmark_index=0,
    model=ModelConfig(
        url="http://vllm-service:8000",
        name="llama-2-7b"
    ),
    parameters={},
    callback_url="http://localhost:8080",
    num_examples=100
)

# Create adapter and run
adapter = ExampleAdapter()
results = adapter.run_benchmark_job(job_spec, callbacks)

Framework Adapter Interface

Your adapter must implement a single method:

from evalhub.adapter import FrameworkAdapter, JobSpec, JobCallbacks, JobResults

class MyFrameworkAdapter(FrameworkAdapter):
    def run_benchmark_job(
        self, config: JobSpec, callbacks: JobCallbacks
    ) -> JobResults:
        """Run a benchmark evaluation job.

        Args:
            config: Job specification from mounted ConfigMap
            callbacks: Callbacks for status updates and artifact persistence

        Returns:
            JobResults: Evaluation results and metadata

        Raises:
            ValueError: If configuration is invalid
            RuntimeError: If evaluation fails
        """
        # Your implementation here
        pass

Key Data Models

JobSpec - Configuration loaded from ConfigMap:

class JobSpec(BaseModel):
    # Mandatory fields
    id: str                           # Unique job identifier
    provider_id: str                   # Provider identifier
    benchmark_id: str                 # Benchmark to evaluate
    benchmark_index: int              # Index of this benchmark within the job (included in all status/result events)
    model: ModelConfig                # Model configuration (url, name)
    parameters: Dict[str, Any]  # Adapter-specific parameters
    callback_url: str                  # Base URL for callbacks (SDK appends /status, /results)

    # Optional fields
    num_examples: Optional[int]       # Number of examples to evaluate
    experiment_name: Optional[str]    # Experiment name
    tags: list[dict[str, str]]        # Custom tags (default: [])

    @classmethod
    def from_file(cls, path: Path | str) -> Self:
        """Load JobSpec from a JSON file."""

Load a job spec from file:

from evalhub.adapter import JobSpec

# Explicit path (recommended)
spec = JobSpec.from_file("/meta/job.json")

# Or use settings for the path
spec = JobSpec.from_file(settings.resolved_job_spec_path)

JobCallbacks - Interface for service communication:

class JobCallbacks(ABC):
    @abstractmethod
    def report_status(self, update: JobStatusUpdate) -> None:
        """Report status update to service"""

    @abstractmethod
    def create_oci_artifact(self, spec: OCIArtifactSpec) -> OCIArtifactResult:
        """Create and push OCI artifact"""

When using DefaultCallbacks, pass benchmark_index (and optionally provider_id) from the job spec so that status and result events sent to the service always include benchmark_index, allowing the service to associate events with the correct benchmark in multi-benchmark jobs.

JobResults - Returned when job completes:

class JobResults(BaseModel):
    id: str
    benchmark_id: str
    benchmark_index: int                       # Index within the job
    model_name: str
    results: List[EvaluationResult]           # Evaluation metrics
    overall_score: Optional[float]            # Overall score if applicable
    num_examples_evaluated: int               # Number of examples evaluated
    duration_seconds: float                   # Total evaluation time
    evaluation_metadata: Dict[str, Any]       # Framework-specific metadata
    oci_artifact: Optional[OCIArtifactResult] # OCI artifact info if persisted
    eval_card: Optional[EvalCardMetadata]     # EvalCard disclosure metadata
    env_card: Optional[EnvironmentCardMetadata] # Environment Card metadata
    additional_info: Optional[Dict[str, Any]]  # Supplementary evaluation info beyond metrics

EvalCard & Environment Card - Evaluation documentation artifacts:

EvalCards and Environment Cards are serialized into the artifacts dict on report_results() and stored by the server — no server changes required.

If a provider does not set env_card, report_results() auto-captures a best-effort Environment Card from the runtime (Python version, OS, GPU info, installed packages). The capture_completeness field (0.0–1.0) reports how many of the 26 spec fields were populated.

# Explicit capture at job start (recommended — captures hardware before eval load)
env_card = EnvironmentCardMetadata.capture(
    framework_name="lm-evaluation-harness",
    framework_version="0.4.5",
)

# EvalCard with capability and safety evaluations
eval_card = EvalCardMetadata(
    modalities_input=["text"],
    modalities_output=["text"],
    languages_count=1,
    languages=["en"],
    capability_evaluations=[
        CapabilityEvalEntry(
            ability="knowledge",
            benchmark="MMLU",
            metric="exact_match",
            alt_prompting=0.712,
            alt_prompting_description="5-Shot",
        ),
    ],
)

# Attach to results before reporting
results = JobResults(..., eval_card=eval_card, env_card=env_card)
callbacks.report_results(results)

additional_info - supplementary evaluation metadata:

additional_info is a dict[str, Any] of supplementary key-value pairs for evaluation information beyond metrics. Values can be any JSON-serializable type, including nested objects and lists. Use additional_info to supply fields such as dataset (list of dataset provenance records), zero_shot, alt_prompting, and alt_prompting_description. It is serialized as a top-level additional_info key in the benchmark_status_event payload and is available to downstream consumers such as EvalCard generation.

Override generate_additional_info() on your FrameworkAdapter subclass to centralise the derivation logic. It is called automatically by DefaultCallbacks.report_results() when results.additional_info is not already set. If a framework has no implementation the base class returns None and it becomes a no-op.

The adapter is not opinionated about where the key-value pairs come from — they can be derived from user input or from the framework's evaluation output. It is up to the implementer to decide. For example, when deriving from lm-evaluation-harness results it could look like:

from evalhub.adapter import (
    FrameworkAdapter,
    JobSpec,
    JobCallbacks,
    JobResults,
)


class LMEvalAdapter(FrameworkAdapter):

    def generate_additional_info(
        self, results: JobResults
    ) -> dict[str, Any] | None:
        """Derive supplementary EvalCard fields from lm-eval output."""
        benchmark_id = results.benchmark_id

        # Resolved n-shot (after task YAML override of the CLI value)
        n_shot = self._n_shot.get(benchmark_id, 0)

        # CoT detection — layered heuristic (no single reliable signal)
        task_config = self._task_configs.get(benchmark_id, {})
        tags = task_config.get("tag", [])
        if isinstance(tags, str):
            tags = [tags]
        doc_to_text = str(task_config.get("doc_to_text", ""))

        is_cot = (
            "chain_of_thought" in tags
            or "cot" in benchmark_id.lower().replace("-", "_").split("_")
            or "think step by step" in doc_to_text.lower()
        )

        is_zero_shot = n_shot == 0 and not is_cot
        score = results.overall_score

        # Build prompting strategy description
        alt_desc = None
        if not is_zero_shot:
            parts = []
            if n_shot > 0:
                parts.append(f"{n_shot}-Shot")
            if is_cot:
                parts.append("CoT")
            alt_desc = " ".join(parts) if parts else None

        return {
            "zero_shot": score if is_zero_shot else None,
            "alt_prompting": score if not is_zero_shot else None,
            "alt_prompting_description": alt_desc,
        }

    def run_benchmark_job(
        self, config: JobSpec, callbacks: JobCallbacks
    ) -> JobResults:
        from lm_eval import simple_evaluate

        lmeval_results = simple_evaluate(...)

        # Store framework output on self for generate_additional_info()
        self._n_shot = lmeval_results.get("n-shot", {})
        self._task_configs = lmeval_results.get("configs", {})

        return JobResults(...)

Deployment

Container Structure

Your adapter runs as a container in a Kubernetes Job alongside a sidecar:

FROM registry.access.redhat.com/ubi9/python-312

WORKDIR /app

# Install your framework and dependencies
RUN pip install lm-evaluation-harness==0.4.0 eval-hub-sdk

# Copy adapter implementation
COPY my_adapter.py .
COPY entrypoint.py .

CMD ["python", "entrypoint.py"]

Entrypoint Script

# entrypoint.py
from my_adapter import MyFrameworkAdapter
from evalhub.adapter import DefaultCallbacks

# Initialize adapter (loads settings and job spec internally)
adapter = MyFrameworkAdapter()

# Create callbacks from adapter (auto-configures sidecar, OCI proxy, etc.)
callbacks = DefaultCallbacks.from_adapter(adapter)

# Run adapter
results = adapter.run_benchmark_job(adapter.job_spec, callbacks)

# Report final results
callbacks.report_results(results)

print(f"Job {results.id} completed with score: {results.overall_score}")

Kubernetes Job

EvalHub creates Jobs automatically:

apiVersion: batch/v1
kind: Job
metadata:
  name: eval-job-123
spec:
  template:
    spec:
      containers:
      - name: adapter
        image: myregistry/my-framework-adapter:latest
        volumeMounts:
        - name: job-spec
          mountPath: /meta
      - name: sidecar
        image: evalhub/sidecar:latest
        env:
        - name: EVALHUB_SERVICE_URL
          value: "http://evalhub-service:8080"
      volumes:
      - name: job-spec
        configMap:
          name: job-123-spec
      restartPolicy: Never

Development

Development Setup

# Clone the repository
git clone https://github.com/eval-hub/eval-hub-sdk.git
cd eval-hub-sdk

# Install in development mode with all dependencies
pip install -e .[dev]

# Install pre-commit hooks
pre-commit install

# Run tests
pytest

# Run tests with coverage
pytest --cov=src/evalhub --cov-report=html

# Run type checking
mypy src/evalhub

# Run linting
ruff check src/ tests/
ruff format src/ tests/

Testing Your Adapter

from evalhub.adapter import AdapterSettings

def test_settings_parse(monkeypatch):
    monkeypatch.setenv("EVALHUB_MODE", "local")
    monkeypatch.setenv("OCI_INSECURE", "true")
    s = AdapterSettings.from_env()
    assert s.oci_insecure is True

Quality Assurance

Run all quality checks:

# Format code
ruff format .

# Lint and fix issues
ruff check --fix .

# Type check
mypy src/evalhub

# Run full test suite
pytest -v --cov=src/evalhub

Installing Pre-Release Versions

Pre-release development versions are published to TestPyPI. To install the latest pre-release:

pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ --pre eval-hub-sdk

The --extra-index-url flag ensures that dependencies are still resolved from the main PyPI index.

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests for your changes
5. Run the test suite
6. Submit a pull request

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.