lm-eval run --model hf --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct --tasks mmlu --output_path ./results/ --log_samples

Writes the evaluation results JSON into ./results/ and, with --log_samples, also saves every individual prompt and model answer for post-hoc analysis. Saved samples let you audit failures and recompute metrics offline. Without --log_samples only aggregated scores are stored.