LLM Learning Hub

workspace/llm-course/home

Benchmarking

Benchmarking is the standardised evaluation of a model on a fixed suite of tasks (MMLU, HumanEval, GSM8K) to produce comparable scores across models and track progress over time.

What is Benchmarking?

Benchmarking is the standardised evaluation of a model on a fixed suite of tasks (MMLU, HumanEval, GSM8K) to produce comparable scores across models and track progress over time.

Benchmarking is the standardised evaluation of a model on a fixed suite of tasks (MMLU, HumanEval, GSM8K) to produce comparable scores across models and track progress over time.

Where is it used?

The Open LLM Leaderboard uses `lm-eval-harness` with MMLU/ARC/HellaSwag/Winogrande; Chatbot Arena uses Elo from human votes; HELM is a comprehensive multi-metric benchmark.

How to build it

Run `lm_eval --model hf --model_args pretrained=<model> --tasks mmlu,arc_challenge,hellaswag,winogrande,gsm8k --output_path results/` and submit to the Open LLM Leaderboard format.