OpenModelsHQ

Evidence

Keep the output. Keep the setup.

A score means more when you can see the task, model revision and configuration behind it. Record a run, compare saved outputs, or choose an evaluation framework for your workload.

Record and inspect outputsSuggest a public evaluation

The recorder works locally in your browser and exports a portable JSON file. It does not call a model or publish your data. OMHQ has not published its own benchmark runs yet; the tools below are external projects, not our test results.

Evaluation tools

Code

EvalPlus

Code-generation evaluation tooling with expanded tests for supported programming benchmarks.

Generated code must run in an isolated environment. Passing a benchmark does not establish safety for arbitrary code execution.

Documentation and source

Agents and language models

Inspect AI

An evaluation framework with tasks, solvers, scorers and logs for language-model and agent evaluations.

Preserve the task definition, model configuration and evaluation log; review tool permissions before running agent tasks.

Documentation and source

Code

LiveCodeBench

A coding benchmark with time-aware problem collection and published evaluation resources.

Match the problem window, dataset version and evaluation setup before comparing reported scores.

Documentation and source

What belongs with a result

Keep the original prompt and output, model identifier and revision, runtime or harness, generation settings and run time. Add observed latency and cost only when measured. A single successful example is not a benchmark, and an exact-text match is not a general quality score.

The harness affects the result. For code or tool-using evaluations, isolate execution and preserve permissions and traces with the result.