lm-eval run --model hf --model_args pretrained=meta-llama/Llama-3.1-8B-Instruct --tasks gsm8k --apply_chat_template

Runs the GSM8K math benchmark with --apply_chat_template so prompts are formatted as chat turns, matching how instruct-tuned models were trained instead of raw text. Skipping this flag tanks scores on chat models. Add --num_fewshot 8 for the standard GSM8K few-shot setup.