detector_benchmark.detector

Submodules

• detector_benchmark.detector.bert_detector
• detector_benchmark.detector.detector
• detector_benchmark.detector.detector_loader
• detector_benchmark.detector.fast_detect_gpt
• detector_benchmark.detector.gpt_zero_detector
• detector_benchmark.detector.watermark_detector

Classes

BertDetector
Detector	Helper class that provides a standard way to create an ABC using
FastDetectGPT
WatermarkDetector
DetectorLoader
GPTZero	Helper class that provides a standard way to create an ABC using

Package Contents

class detector_benchmark.detector.BertDetector(model: torch.nn.Module, tokenizer: transformers.PreTrainedTokenizerBase, device: str)

Bases: detector_benchmark.detector.detector.Detector

model

tokenizer

device

detect(texts: list[str], batch_size: int, detection_threshold: float = 0.0) → tuple[list[int], list[float], list[int]]

Detect the if the texts given as input are AI-generated (label 1) or human-written (label 0). Returns the predicted lables with argmax, the logits of the positive class and the predicted labels with the given detection threshold instead of the argmax.

Parameters:

texts: list[str]

The texts to detect

batch_size: int

The batch size to use for detection

detection_threshold: float

The threshold to use for detection. Default is 0.0.

Returns:

tuple[list[int], list[float], list[int]]

The predicted labels with argmax, the logits of the positive class and the predicted labels with the given detection threshold instead of the argmax.

class detector_benchmark.detector.Detector

Bases: abc.ABC

Helper class that provides a standard way to create an ABC using inheritance.

abstractmethod detect(texts: list[str], batch_size: int, detection_threshold: float = 0.0) → tuple[list[int], list[float], list[int]]

Detect the watermark in the texts.

Parameters:

texts: list[str]

The texts to detect the watermark in

batch_size: int

The batch size

detection_threshold: float

The threshold to use for the detection

Returns:

tuple[list[int], list[float], list[int]]

The predictions, the logits, and the predictions at the threshold

class detector_benchmark.detector.FastDetectGPT(ref_model, scoring_model, ref_tokenizer, scoring_tokenizer, device)

Bases: detector_benchmark.detector.detector.Detector

ref_model

scoring_model

ref_tokenizer

scoring_tokenizer

device

get_samples(logits, labels) → torch.Tensor

Get the samples from the logits.

Parameters:

logits: torch.Tensor

The logits

labels: torch.Tensor

The labels

get_likelihood(logits, labels) → torch.Tensor

Get the likelihood from the logits.

Parameters:

logits: torch.Tensor

The logits

labels: torch.Tensor

The labels

get_sampling_discrepancy(logits_ref, logits_score, labels) → torch.Tensor

Get the sampling discrepancy from the logits.

Parameters:

logits_ref: torch.Tensor

The logits of the reference model

logits_score: torch.Tensor

The logits of the scoring model

labels: torch.Tensor

The labels

get_sampling_discrepancy_analytic(logits_ref, logits_score, labels) → torch.Tensor

Get the sampling discrepancy from the logits.

Parameters:

logits_ref: torch.Tensor

The logits of the reference model

logits_score: torch.Tensor

The logits of the scoring model

class ProbEstimatorFastDetectGPT(args=None, ref_path=None)

Probability estimator for the FastDetectGPT detector.

real_crits = []

fake_crits = []

crit_to_prob(crit) → float

Convert the criterion to probability.

Parameters:

crit: float

The criterion

Returns:

float

The probability

detect(texts: list[str], batch_size: int, detection_threshold: float = 0.5) → tuple[list[int], list[float], list[int]]

Detect the watermark in the texts.

Parameters:

texts: list[str]

The texts to detect the watermark in

batch_size: int

The batch size

detection_threshold: float

The detection threshold

Returns:

tuple[list[int], list[float], list[int]]

The predictions, the probabilities and the predictions at the threshold

class detector_benchmark.detector.WatermarkDetector(watermarking_scheme: detector_benchmark.watermark.auto_watermark.AutoWatermark, detection_threshold: float)

Bases: detector_benchmark.detector.detector.Detector

watermarking_scheme

detection_threshold

detect(texts: list[str], batch_size: int, detection_threshold: float) → tuple[list[int], list[float], list[int]]

Detect the if the texts given as input are watermarked (label 1) or not (label 0).

Parameters:

texts: list[str]

The texts to detect

batch_size: int

The batch size

detection_threshold: float

The threshold to use for the detection

Returns:

tuple[list[int], list[float], list[int]]

The predictions, the logits for the positive class, and the predictions at the threshold

class detector_benchmark.detector.DetectorLoader(cfg: dict, detector_name: str, device: str, weights_checkpoint: str = None, local_weights: bool = False)

cfg

detector_name

device

weights_checkpoint = None

local_weights = False

load() → detector_benchmark.detector.detector.Detector

Load the detector based on the given configuration (init).

Returns:

Detector

The loaded detector

class detector_benchmark.detector.GPTZero(api_key, debug_mode=False)

Bases: detector_benchmark.detector.detector.Detector

Helper class that provides a standard way to create an ABC using inheritance.

api_key

debug_mode = False

predict_gpt_zero(text, api_key, debug_mode=False) → dict

Predict the GPT-Zero score for the text.

Parameters:

text: str

The text to predict

api_key: str

The API key

debug_mode: bool

Whether to print the debug information

Returns:

dict

The prediction result

detect(texts: list, batch_size: int, detection_threshold: float = 0.5) → tuple[list[int], list[float], list[int]]

Detect the GPT-Zero score for the texts.

Parameters:

texts: list

The texts to detect

batch_size: int

The batch size

detection_threshold: float

The threshold to use for the detection

Returns:

tuple[list[int], list[float], list[int]]

The predictions, the logits for the positive class, and the predictions at the threshold