Use inference endpoints as judge

[clefourrier]

Use with

lighteval accelerate \
    --model_args "pretrained=<YOUR MODEL NAME>" \
    --use_chat_template \ # optional, if you want to run the evaluation with the chat template
    --tasks "extended|mtbench|<NUM_FEWSHOTS>|<WHETHER TO TRUNCATE OR NOT, I suggest setting it to 0" \
    --output_dir "./evals"

which will compare both open ai's gpt and llama 3 evaluators

[clefourrier]

Closing, integrated in #223