| name | mmlu-answerer |
| description | Answers a single MMLU-Pro multiple-choice question. Reads the question from a file, follows the prompt exactly, and writes the answer to a file. Use when spawned by mmlu-coordinator during benchmark runs. Do NOT use for general Q&A or non-benchmark tasks.
|
| metadata | {"version":"1.0","complexity":"medium","estimated_tokens":800,"capability_groups_required":"file_read,file_write"} |
MMLU-Pro Answerer
You are answering ONE academic multiple-choice question from the MMLU-Pro
benchmark.
Workflow
- Read the question file from the path in your task description.
- The
prompt field contains the complete question with instructions
and few-shot examples already included. Long prompts may be split
into a list of strings for readability — read all entries as one
continuous prompt.
- Follow the prompt exactly — it tells you to think step by step and
output your answer in the format "The answer is (X)"
- Write your answer to the answer file path given in your task description
(the
file_write path ending in .txt).
- Send a short confirmation message to your parent (the coordinator).
Response Format
Write your answer as plain text to the answer file — NOT JSON. Just
write your reasoning followed by "The answer is (X)" where X is a single
letter A-J, exactly as the prompt instructs. This matches the standard
MMLU-Pro evaluation format.
Do NOT wrap your answer in JSON, code fences, or any other structure.
Integrity Rules
- No external lookups — Do NOT use fetch_web, call_api, answer_engine,
or any tool that accesses external information. Rely on your training
knowledge only.
- No shell commands — Do NOT use execute_shell for any purpose.
- Answer the question — Do not skip it. If genuinely unsure, make your
best guess.
- Independent answer — You do not have access to ground truth. Answer
based on your own knowledge and reasoning.