eval-hub-contrib

Community-contributed evaluation framework adapters for eval-hub.

Overview

This repository contains adapters that integrate various evaluation frameworks with the eval-hub service. Each adapter implements the FrameworkAdapter pattern from the evalhub-sdk, enabling seamless integration with the eval-hub evaluation service.

Supported Frameworks

Framework	Container Image	Kubernetes	Notes
LightEval	quay.io/evalhub/community-lighteval:latest	✓	Lightweight evaluation framework for language models
GuideLLM	quay.io/evalhub/community-guidellm:latest	✓	Performance benchmarking for LLM inference servers
MTEB	quay.io/evalhub/community-mteb:latest	✓	Massive Text Embedding Benchmark for embedding models
IBM CLEAR	quay.io/evalhub/community-ibm-clear:latest	✓	Agentic trace analysis (LLM-as-judge error reporting)
Inspect AI	quay.io/evalhub/community-inspect:latest	✓	UK AISI framework — 75 benchmarks total: 36 Petri alignment audits, 2 Bloom suites, and 37 inspect-evals benchmarks
RAGAS	quay.io/evalhub/community-ragas:latest	✓	RAG pipeline quality evaluation (faithfulness, relevancy, context precision/recall, and more)
SWE-bench	quay.io/evalhub/community-swebench:latest	✓	Software engineering benchmark for code patch evaluation
DeepEval	quay.io/evalhub/community-deepeval:latest	✓	LLM-as-judge evaluation: faithfulness, relevancy, hallucination, correctness, summarization, and multi-turn conversation metrics
RULER	quay.io/evalhub/community-ruler:latest	✓	NVIDIA RULER long-context benchmark — 13 synthetic tasks across needle-in-a-haystack, variable tracking, aggregation, and QA at configurable context lengths
WildGuard	quay.io/evalhub/community-wildguard:latest	✓	AllenAI safety classification benchmark — evaluates a model's ability to classify prompt+response pairs as safe or unsafe, reporting accuracy and per-class recall

Inspect AI Adapter

The Inspect AI adapter exposes alignment auditing and safety evaluation through the Petri and Bloom tools from Meridian Labs, as well as curated benchmarks from the inspect-evals community library.

75 benchmarks across three categories:

• 36 Petri alignment audits — covers all 40 built-in seed tag categories including sycophancy, deception, alignment faking, jailbreak, harmful cooperation, self-preservation, power seeking, oversight subversion, and more.
• 2 Bloom behavioral suites — automated scenario generation from high-level behavior descriptions.
• 37 inspect-evals — safety (AgentHarm, WMDP, StrongREJECT, MASK), scheming (agentic misalignment, GDM self-proliferation, GDM stealth), cybersecurity (Cybench, CyberSecEval), coding (HumanEval, SWE-bench), math (GSM8K, MATH, AIME), knowledge (MMLU, GPQA), and agent capabilities (GAIA, TheAgentCompany).

Model configuration — no provider prefixes required in job specs. The adapter detects the correct API from environment variables:

Environment variable	API used
OPENAI_BASE_URL + OPENAI_API_KEY	OpenAI-compatible (vLLM, OpenRouter)
OLLAMA_BASE_URL or port 11434	Ollama native
ANTHROPIC_API_KEY	Anthropic Messages API

See adapters/inspect/README.md for full documentation, deployment examples, and benchmark catalog.

DeepEval Adapter

The DeepEval adapter integrates DeepEval into eval-hub using an LLM-as-judge approach. A separate judge model scores outputs against configurable thresholds. Test data is loaded from CSV, JSONL, or JSON files and mapped to either single-turn or multi-turn DeepEval test cases.

8 benchmarks across two categories:

• 5 single-turn — faithfulness (retrieval grounding), answer relevancy, hallucination detection, factual correctness, and summarization quality.
• 3 multi-turn — conversation completeness (all user needs are addressed), role adherence (chatbot stays in persona), and knowledge retention (chatbot recalls user-disclosed information across turns).

Judge model configuration — the adapter accepts an independent judge model separate from the evaluated model:

Parameter	Description
eval_model_name	Judge model name (defaults to the evaluated model)
eval_model_url	OpenAI-compatible base URL for the judge endpoint
threshold	Minimum pass score (default 0.5)
dataset_format	Input format: csv, jsonl, or json (default csv)

See adapters/deepeval/README.md for full documentation, dataset column requirements, and multi-turn conversation format.

RULER Adapter

The RULER adapter integrates NVIDIA RULER (What's the Real Context Size of Your LLM?) into eval-hub. RULER is a synthetic long-context benchmark that evaluates LLMs across 13 tasks at configurable context lengths, measuring effective context utilisation rather than a simple token-count ceiling.

13 benchmarks across four categories:

• 8 Needle-in-a-Haystack — single needle with noise/essay/UUID haystacks, multi-key, multi-value, multi-query, and needle-background variants.
• 1 Variable Tracking — track chains of variable assignments and return the final value.
• 2 Aggregation — identify the most frequent words and the top coded words in a Zipf-distributed list.
• 2 Question Answering — long-context QA with SQuAD and HotpotQA passages.

Model configuration — any OpenAI-compatible inference endpoint. Set model.url to the /v1 endpoint and model.name to the model ID.

Key parameters:

Parameter	Default	Description
context_lengths	[4096, 8192, 16384]	Context sizes (tokens) to evaluate
num_samples	10	Samples per (task × context length)
tokenizer_path	model.name	HuggingFace model ID or tiktoken model for data generation
tokenizer_type	hf	hf (HuggingFace) or openai (tiktoken)
model_template	base	Chat prompt template (see scripts/data/template.py)
random_seed	42	Seed for reproducible data generation

See adapters/ruler/README.md for full documentation, example job specs, and vendored script details.

WildGuard Adapter

The WildGuard adapter integrates the WildGuard safety benchmark (allenai/wildguard, MIT licence) from AllenAI. For each prompt+response pair in the dataset, the adapter sends the WildGuard instruction template to the model, parses its natural-language output as safe or unsafe (outputs containing neither are treated as unknown), and compares against the ground-truth label. Unknown predictions are counted as incorrect when calculating accuracy.

1 benchmark:

• wildguard-safety — evaluates the model against the WildGuard test split, reporting accuracy and per-class recall.

Metrics:

Metric	Description
accuracy	Fraction of examples correctly classified (overall_score)
safe_recall	Recall on genuinely safe responses
unsafe_recall	Recall on genuinely unsafe responses

Score guide: 0.5 or below is near random chance; 0.85 or above is strong performance. Low unsafe_recall indicates under-refusal; low safe_recall indicates over-refusal.

Key parameters:

Parameter	Default	Description
split	test	HuggingFace dataset split
num_examples	(full split)	Cap the number of examples (useful for smoke tests)
max_concurrent	4	Concurrent API calls to the model endpoint
request_timeout	120	Per-request timeout in seconds

See adapters/wildguard/README.md for full documentation and example job specs.

JobPhase Lifecycle

Every adapter must report progress through the JobPhase lifecycle via callbacks.report_status(). The server validates phases against a fixed set, so adapters must emit them in order and use only the values listed below.

Phases

1. INITIALIZING — Validate configuration, resolve credentials, set up temporary directories. Emit at the start of run_benchmark_job.
2. LOADING_DATA — Load datasets, download test data, prepare inputs. Emit before any data I/O.
3. RUNNING_EVALUATION — Execute the framework (subprocess, API call, etc.). Emit before the main workload begins.
4. POST_PROCESSING — Parse results, extract metrics, compute scores. Emit after the framework finishes.
5. PERSISTING_ARTIFACTS — Create OCI artifacts from result files. Emit only when OCI exports are configured (config.exports.oci). Skip this phase entirely when there is nothing to persist.
6. COMPLETED — Do not emit manually. This phase is sent automatically by callbacks.report_results().

Status update format

Only status and phase are forwarded to the server. Other fields (progress, message, current_step, etc.) are silently dropped by the SDK.

# Success path — emit for each phase
callbacks.report_status(
    JobStatusUpdate(status=JobStatus.RUNNING, phase=JobPhase.INITIALIZING)
)

# Failure path — use error_message (ErrorInfo is deprecated)
callbacks.report_status(
    JobStatusUpdate(
        status=JobStatus.FAILED,
        error_message=MessageInfo(message=str(e), message_code="evaluation_error"),
    )
)

PERSISTING_ARTIFACTS gating

The PERSISTING_ARTIFACTS phase must only be reported when OCI exports are configured. When no exports are configured, skip both the phase and the OCI call:

oci_artifact = None
oci_exports = config.exports.oci if config.exports else None
if oci_exports is not None and output_files:
    callbacks.report_status(
        JobStatusUpdate(status=JobStatus.RUNNING, phase=JobPhase.PERSISTING_ARTIFACTS)
    )
    oci_artifact = callbacks.create_oci_artifact(
        OCIArtifactSpec(files_path=results_dir, coordinates=oci_exports.coordinates)
    )

Building Adapters

# Build a specific adapter
make image-lighteval
make image-guidellm
make image-mteb
make image-inspect
make image-deepeval
make image-ragas
make image-swebench

# Build all adapters
make images

# Run adapter tests
make test-lighteval
make test-inspect
make test-deepeval
make test-ragas
make test-swebench
make test-clear
make tests

# Push to registry
make push-lighteval REGISTRY=quay.io/your-org VERSION=v1.0.0
make push-mteb REGISTRY=quay.io/your-org VERSION=v1.0.0
make push-inspect REGISTRY=quay.io/your-org VERSION=v1.0.0
make push-deepeval REGISTRY=quay.io/your-org VERSION=v1.0.0
make push-ragas REGISTRY=quay.io/your-org VERSION=v1.0.0
make push-swebench REGISTRY=quay.io/your-org VERSION=v1.0.0

Contributing

See CONTRIBUTING.md for guidelines on adding adapters.

License

See the LICENSE file for details.