Possible mistake in the evaluation of IFEval
Hello, I am writing because I noticed that the results for the IFEval task seem to be excessively low compared to the previous evaluation cycle. Did something go wrong in the latest evaluation of this model?
I have also evaluated it independently by using your code, and this is what I get:
"results": {
"leaderboard_ifeval": {
"alias": "leaderboard_ifeval",
"prompt_level_strict_acc,none": 0.7338262476894639,
"prompt_level_strict_acc_stderr,none": 0.01901876684729058,
"inst_level_strict_acc,none": 0.8177458033573142,
"inst_level_strict_acc_stderr,none": "N/A",
"prompt_level_loose_acc,none": 0.7763401109057301,
"prompt_level_loose_acc_stderr,none": 0.017931771054658374,
"inst_level_loose_acc,none": 0.8465227817745803,
"inst_level_loose_acc_stderr,none": "N/A"
}
I am pretty sure this is similar to how it looked like in the leaderboard previously.
This is the command I used in input:
lm-eval --model_args="pretrained=meta-llama/Llama-3.1-8B-Instruct,revision=main,dtype=bfloat16" --tasks=leaderboard_ifeval --batch_size=auto --output_path=results --apply_chat_template --fewshot_as_multiturn
Thank you in advance for the clarification!