Implementation:Liu00222 Open Prompt Injection compute conditional probability

Knowledge Sources	Open-Prompt-Injection
Domains	NLP, Language_Modeling
Last Updated	2026-02-14 15:00 GMT

Overview

Concrete function for computing conditional log-probabilities of target text given a prefix using GPT-2, provided by the PromptLocate module.

Description

The compute_conditional_probability function concatenates the condition and target text, tokenizes them, runs a forward pass through the GPT-2 model, and extracts log-probabilities for only the target token positions. It handles truncation to 1024 tokens (GPT-2 maximum context) and returns both average and total log-probability.

Usage

Called by `causal_influence` to compute `P(suffix|prefix)` and `P(suffix|prefix+injected)`. Not typically called directly by users.

Code Reference

Source Location

• Repository: Open-Prompt-Injection
• File: OpenPromptInjection/apps/PromptLocate.py
• Lines: L123-151

Signature

def compute_conditional_probability(condition_text, target_text, tokenizer, model):
    """
    Compute conditional log-probability of target_text given condition_text.

    Args:
        condition_text (str): Conditioning prefix text.
        target_text (str): Target text to compute probability for.
        tokenizer: GPT-2 tokenizer.
        model: GPT-2 model on CUDA.
    Returns:
        tuple[float, float]: (avg_log_prob, total_log_prob).
            avg_log_prob: Average log-probability per target token.
            total_log_prob: Sum of log-probabilities for all target tokens.
            Truncates combined input to 1024 tokens max.
    """

Import

from OpenPromptInjection.apps.PromptLocate import compute_conditional_probability

I/O Contract

Inputs

Name	Type	Required	Description
condition_text	str	Yes	Conditioning prefix text
target_text	str	Yes	Target text to compute probability for
tokenizer	PreTrainedTokenizer	Yes	GPT-2 tokenizer
model	PreTrainedModel	Yes	GPT-2 model on CUDA

Outputs

Name	Type	Description
avg_log_prob	float	Average log-probability per target token (more negative = less likely)
total_log_prob	float	Sum of log-probabilities for all target tokens

Usage Examples

Computing Conditional Probability

from OpenPromptInjection.apps.PromptLocate import compute_conditional_probability
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("gpt2")
model = AutoModelForCausalLM.from_pretrained("gpt2").cuda()

condition = "The weather is sunny today."
target = " The temperature is 75 degrees."

avg_logp, total_logp = compute_conditional_probability(condition, target, tokenizer, model)
print(f"Avg log-prob: {avg_logp:.4f}")
print(f"Total log-prob: {total_logp:.4f}")

Related Pages

Implements Principle

• Principle:Liu00222_Open_Prompt_Injection_Conditional_Probability_Computation

Requires Environment

• Environment:Liu00222_Open_Prompt_Injection_CUDA_Environment