KomeijiForce/roberta-large-metaie-gpt4

MetaIE

This is a meta-model distilled from ChatGPT-4 for information extraction. This is an intermediate checkpoint that can be well-transferred to all kinds of downstream information extraction tasks. This model can also be tested by different label-to-span matching as shown in the following example:

from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch

device = torch.device("cuda:0")
path = f"KomeijiForce/roberta-large-metaie-gpt4"
tokenizer = AutoTokenizer.from_pretrained(path)
tagger = AutoModelForTokenClassification.from_pretrained(path).to(device)

def find_sequences(lst):
    sequences = []
    i = 0
    while i < len(lst):
        if lst[i] == 0:
            start = i
            end = i
            i += 1
            while i < len(lst) and lst[i] == 1:
                end = i
                i += 1
            sequences.append((start, end+1))
        else:
            i += 1
    return sequences

def is_sublst(lst1, lst2):
    for idx in range(len(lst1)-len(lst2)+1):
        if lst1[idx:idx+len(lst2)] == lst2:
            return True
    return False

words = ["John", "Smith", "loves", "his", "hometown", ",", "Los", "Angeles", "."]

for prefix in ["Person", "Location", "John Smith births in", "Positive opinion"]:

    sentence = " ".join([prefix, ":"]+words)

    inputs = tokenizer(sentence, return_tensors="pt").to(device)
    tag_predictions = tagger(**inputs).logits[0].argmax(-1)

    predictions = [tokenizer.decode(inputs.input_ids[0, seq[0]:seq[1]]).strip() for seq in find_sequences(tag_predictions)]
    predictions = [prediction for prediction in predictions if is_sublst(words, prediction.split())]
    
    print(prefix, predictions)

The output will be

"Person" ['John Smith']
"Location" ['Los Angeles']
"John Smith births in" ['Los Angeles']
"Positive opinion" ['loves his hometown']