potsawee/mmlu-llm-zero-shot

HaluEval

• Data: https://huggingface.co/datasets/cais/mmlu and using load_dataset("cais/mmlu", "all", split='test')
• To obtain the original question and options:

from datasets import load_dataset
data = load_dataset("cais/mmlu", "all", split='test') # the size should match the size of the data in this repo

• Judge prompt: similar to https://www.kaggle.com/code/debarshichanda/llm-evaluation-mmlu-style
• size: 14042

This repository

• mmlu-llm-zero-shot folder contains predictions made by each judge (logits over A, B, C, and D)
• pred and label: 0 -> "A", 1 -> "B", 2 -> "C", 3 -> "D"

LLM Judge Statistics

Judge	Accuracy
RANDOM GUESSING	25.00
dolphin-2.1-mistral-7b	35.13
StableBeluga-7B	50.57
Meta-Llama-3-8B	46.09
Mistral-7B-Instruct-v0.1	51.18
Mistral-7B-Instruct-v0.2	56.32
zephyr-7b-beta	56.88
Mistral-7B-OpenOrca	57.93
Meta-Llama-3-8B-Instruct	44.03
OpenHermes-2-Mistral-7B	57.78
OpenHermes-2.5-Mistral-7B	58.94
Starling-LM-7B-alpha	59.77
:---------------------------------	:-------
Meta-Llama-3-70B-Instruct	77.76
Mixtral-8x7B-Instruct-v0.1	65.82
Qwen2-72B-Instruct	79.02
Hermes-3-Llama-3.1-70B	75.70
Athene-70B	76.71
dolphin-2.5-mixtral-8x7b	48.20

Example of Judge's prompt

system_prompt = """You are a helpful assistant and your task is to answer the following multiple-choice question by giving the most appropriate response. Answer should be one among [A, B, C, D]. The response must stricly follow this format: Answer: [[x]] where x is the letter answer such as A, B, C, D."""

prompt_template = """Question: {question}\n
A) {a}\n
B) {b}\n
C) {c}\n
D) {d}\n

Answer: [["""

messages=[
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt}, # I just realised it might be better to add another turn for "assistant"
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False) # then remove all special tokens after `Judgement#: [[`