app.py

import torch
import joblib

import numpy as np
import pandas as pd
import gradio as gr

from nltk.data import load as nltk_load
from transformers import AutoTokenizer, AutoModelForCausalLM


print("Loading model & Tokenizer...")
model_id  = 'gpt2'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForCausalLM.from_pretrained(model_id)

print("Loading NLTL & and scikit-learn model...")
NLTK = nltk_load('data/english.pickle')
sent_cut_en = NLTK.tokenize
clf = joblib.load(f'data/gpt2-small-model')

CROSS_ENTROPY = torch.nn.CrossEntropyLoss(reduction='none')

example = """\
The perplexity (PPL) is commonly used as a metric for evaluating the performance of language models (LM). It is defined as the \
exponential of the negative average log-likelihood of the text under the LM. A lower PPL indicates that the language model is more confident \
in its predictions, and is therefore considered to be a better model. The training of LMs is carried out on large-scale text corpora, it can \
be considered that it has learned some common language patterns and text structures. Therefore, PPL can be used to measure how well a text \
conforms to common characteristics.

I used all variants of the open-source GPT-2 model except xl size to compute the PPL (both text-level and sentence-level PPLs) of the collected \
texts. It is observed that, regardless of whether it is at the text level or the sentence level, the content generated by LLMs have relatively \
lower PPLs compared to the text written by humans. LLM captured common patterns and structures in the text it was trained on, and is very good at \
reproducing them. As a result, text generated by LLMs have relatively concentrated low PPLs.\
"""


def gpt2_features(text, tokenizer, model, sent_cut):
    # Tokenize
    input_max_length = tokenizer.model_max_length - 2
    token_ids, offsets = list(), list()
    sentences = sent_cut(text)
    for s in sentences:
        tokens = tokenizer.tokenize(s)
        ids = tokenizer.convert_tokens_to_ids(tokens)
        difference = len(token_ids) + len(ids) - input_max_length
        if difference > 0:
            ids = ids[:-difference]
        offsets.append((len(token_ids), len(token_ids) + len(ids)))
        token_ids.extend(ids)
        if difference >= 0:
            break

    input_ids = torch.tensor([tokenizer.bos_token_id] + token_ids)
    logits = model(input_ids).logits
    # Shift so that n-1 predict n
    shift_logits = logits[:-1].contiguous()
    shift_target = input_ids[1:].contiguous()
    loss = CROSS_ENTROPY(shift_logits, shift_target)

    all_probs = torch.softmax(shift_logits, dim=-1)
    sorted_ids = torch.argsort(all_probs, dim=-1, descending=True)  # stable=True
    expanded_tokens = shift_target.unsqueeze(-1).expand_as(sorted_ids)
    indices = torch.where(sorted_ids == expanded_tokens)
    rank = indices[-1]
    counter = [
        rank < 10,
        (rank >= 10) & (rank < 100),
        (rank >= 100) & (rank < 1000),
        rank >= 1000
    ]
    counter = [c.long().sum(-1).item() for c in counter]


    # compute different-level ppl
    text_ppl = loss.mean().exp().item()
    sent_ppl = list()
    for start, end in offsets:
        nll = loss[start: end].sum() / (end - start)
        sent_ppl.append(nll.exp().item())
    max_sent_ppl = max(sent_ppl)
    sent_ppl_avg = sum(sent_ppl) / len(sent_ppl)
    if len(sent_ppl) > 1:
        sent_ppl_std = torch.std(torch.tensor(sent_ppl)).item()
    else:
        sent_ppl_std = 0

    mask = torch.tensor([1] * loss.size(0))
    step_ppl = loss.cumsum(dim=-1).div(mask.cumsum(dim=-1)).exp()
    max_step_ppl = step_ppl.max(dim=-1)[0].item()
    step_ppl_avg = step_ppl.sum(dim=-1).div(loss.size(0)).item()
    if step_ppl.size(0) > 1:
        step_ppl_std = step_ppl.std().item()
    else:
        step_ppl_std = 0
    ppls = [
        text_ppl, max_sent_ppl, sent_ppl_avg, sent_ppl_std,
        max_step_ppl, step_ppl_avg, step_ppl_std
    ]
    return ppls + counter  # type: ignore


def predict_out(features, classifier, id_to_label):
    x = np.asarray([features])
    pred = classifier.predict(x)[0]
    prob = classifier.predict_proba(x)[0, pred]
    return [id_to_label[pred], prob]


def predict(text):
    with torch.no_grad():
        feats = gpt2_features(text, tokenizer, model, sent_cut_en)
    out = predict_out(feats, clf, ['Human Written', 'LLM Generated'])
    return out


with gr.Blocks() as demo:
    gr.Markdown(
        """\
        ## Detect text generated using LLMs 🤖

        Linguistic features such as Perplexity and other SOTA methods such as GLTR were used to classify between Human written and LLM Generated \
        texts. This solution scored an ROC of 0.956 and 8th position in the DAIGT LLM Competition on Kaggle.

        - Source & Credits: [https://github.com/Hello-SimpleAI/chatgpt-comparison-detection](https://github.com/Hello-SimpleAI/chatgpt-comparison-detection)
        - Competition: [https://www.kaggle.com/competitions/llm-detect-ai-generated-text/leaderboard](https://www.kaggle.com/competitions/llm-detect-ai-generated-text/leaderboard)
        - Solution WriteUp: [https://www.kaggle.com/competitions/llm-detect-ai-generated-text/discussion/470224](https://www.kaggle.com/competitions/llm-detect-ai-generated-text/discussion/470224)\
        """
    )
    with gr.Row():
        gr.Markdown(
            """\
            ### Linguistic Analysis: Language Model Perplexity
            The perplexity (PPL) is commonly used as a metric for evaluating the performance of language models (LM). It is defined as the exponential \
            of the negative average log-likelihood of the text under the LM. A lower PPL indicates that the language model is more confident in its \
            predictions, and is therefore considered to be a better model. The training of LMs is carried out on large-scale text corpora, it can \
            be considered that it has learned some common language patterns and text structures. Therefore, PPL can be used to measure how \
            well a text conforms to common characteristics.
    
            ### GLTR: Giant Language Model Test Room
            This idea originates from the following paper: arxiv.org/pdf/1906.04043.pdf. It studies 3 tests to compute features of an input text. Their \
            major assumption is that to generate fluent and natural-looking text, most decoding strategies sample high probability tokens from the head \
            of the distribution. I selected the most powerful Test-2 feature, which is the number of tokens in the Top-10, Top-100, Top-1000, and 1000+ \
            ranks from the LM predicted probability distributions.
    
            ### Modelling
            Scikit-learn's VotingClassifier consisting of XGBClassifier, LGBMClassifier, CatBoostClassifier and RandomForestClassifier with default parameters\
            """
        )
        with gr.Column():
            a1 = gr.Textbox( lines=7, label='Text', value=example )
            button1 = gr.Button("🤖 Predict!")
            gr.Markdown("Prediction:")
            label1 = gr.Textbox(lines=1, label='Predicted Label')
            score1 = gr.Textbox(lines=1, label='Predicted Probability')
        
            button1.click(predict, inputs=[a1], outputs=[label1, score1])

demo.launch()