[EVAL] Aligning with oz-eval benchmark

[djmilena]

Aligning with oz-eval benchmark

oz-eval is evaluated through the pattern:

Pitanje: {question}

Ponuđeni odgovori:
A. {option_a}
B. {option_b}
C. {option_c}
D. {option_d}
E. {option_e}

Krajnji odgovor:

Please note that in your serbian_eval.py there is

 instruction = "Na osnovu sledećeg pitanja, izaberite tačanu opciju iz ponuđenih odgovora.\n"

which we are not using and should be avoided when evaluating this dataset.
In addition, this enumeration is not in line as well:

   for i, choice in enumerate(choices):
        query += f"{i}. {choice}\n"

since in our case array of letters A-E as shown in the example above.

Evaluation metadata

• GitHub url: https://github.com/Stopwolf/lighteval/blob/oz-eval/community_tasks/oz_evals.py
• Dataset url: https://huggingface.co/datasets/DjMel/oz-eval

Hope this helps make the changes! We are available for any open question and further info 🤗

[DeanChugall]

@djmilena
should be fixed now.

Can you confirm that is good!?