ML-Patch: Carefully Evaluating Hidden Knowledge of Language Models via Multi-layer Patching
We present ML-Patch, a new evaluation method of LLMs, which consists of an online website to show our method clealy and an easy-use online inference website which can infer the result of ML-Patch online using user's own data. Moreover, we also provide an offline python toolkit for users who want to upload large amounts of data.
Specifically, we propose a new method to evaluate the knowledge boundry pf LLMs, which can make better use of the hidden states of LLMs. It is significantly different from today's evaluation methods which most base on prompt.
We test our method on a series of pretrained models, including llama2-13b ,gpt-j-6b, Qwen2.5-7b e.t.c. The results show that our method can effectively evaluate the knowledge of LLMs.
We use the factual triples sorted out from wikidata.
By running api.ipynb, you can input the factual knowledge and choose a series of hyperparameters such as model and get a pkl and tsv file which contain the final results.
import pandas as pd
import io
import os
from ML_patch import *
from zhipuai import ZhipuAI
import requests
client = ZhipuAI(api_key="")
# Here we use ZhipuAI to generate a sentence which contains the subject
# Please use your own api key here.
data_ = "id,subject,relation,object\n001,France,capital city of,Paris"
bytes_io = io.StringIO(data_)
df = pd.read_csv(bytes_io, sep=",") # You can load your own data here
result = Ml_patch(model_name= "/data3/MODELS/gpt-j-6b" , data = df, only_final_result= False, patch_num= 3,client=client)
result.to_csv("./patch.tsv", sep="\t", index=False)
result.to_pkl("./patch.pkl")
More hyperparameters can be adjusted in ML_Patch.py