Using the following input/output formats to achieve the best results.

from transformers import AutoTokenizer
from transformers import AutoModelForSequenceClassification
import torch

tokenizer=AutoTokenizer.from_pretrain("SunW7777/EpicPRM")
model=AutoModelForSequenceClassification.from_pretrain("SunW7777/EpicPRM")
prompt = "### Instruct:\nThe steps in 'Selected steps' are the correct problem-solving steps of the problem, while the steps in 'Next step:' are the next problem-solving steps generated by an AI agent based on the steps in 'Selected steps:'.You need to rate the step in 'Next step:' based on it`s usefulness and correctness.\n\n"
state="Problem:There are 3 complex numbers $a+bi$, $c+di$, and $e+fi$. If $b=3$, $e=-a-c$, and the sum of the numbers is $2i$, find $d+f$. Selected steps: Step 1: We are given three complex numbers in the form $a+bi$, $c+di$, and $e+fi$. Step 2: We're also given that $b=3$ and $e=-a-c$."
next_step="Step 3: The sum of the three numbers is given as $2i$, which means $(a+bi) + (c+di) + (e+fi) = 2i$.\n"
input_text=prompt+state+next_step
tokenized_input = tokenizer(input_text, return_tensors="pt", max_length=8000, truncation=True).input_ids[0]
print(model(**tokenized_input))

Train dataset is here (https://huggingface.co./datasets/SunW7777/EpicPRM)