EleutherAI/lm-evaluation-harness ? reverse-engineered prompt

Build me a Python tool that can run repeatable evaluations for language models on a bunch of common benchmark tasks.

I want to be able to pick a model, pick one or more tasks, run everything from the command line, and get a clean results summary with scores and logs. It should support local Hugging Face models, faster local inference, and API based models too, and it should be easy to add custom tasks, prompts, and metrics without changing the core code. A config file option would be great so I can save and reuse setups instead of typing long commands every time.

Please include a simple way to list available tasks, run evaluations, and validate configs. Make the project easy to install, with optional extras for different model backends, and keep the code organized so tests and docs can grow with it. If you need to check current docs or package behavior online, do that.