get_prompts
Returns text prompts for a specific dataset and guidance type.
from laft.prompts import get_prompts
prompts = get_prompts(
dataset_name="color_mnist",
guidance="guide_number"
)
Parameters
Name of the dataset to get prompts for.Supported datasets:
color_mnist: Colored MNIST digitswaterbirds: Waterbirds datasetceleba: CelebA dataset
Type of guidance to use. Can be prefixed with guide_ or ignore_.For color_mnist:
guide_number or number: Guide attention to digit numbersguide_color or color: Guide attention to colorsignore_number: Ignore number variationsignore_color: Ignore color variations
The function automatically adds guide_ prefix if not present. Returns
Dictionary mapping prompt categories to lists of text prompts.Keys:
normal: Prompts for normal/expected attributesanomaly: Prompts for anomalous attributeshalf: Prompts for normal + half of anomaly attributesexact: Prompts for normal + all anomaly attributesall: Prompts for normal + anomaly + auxiliary attributes
For guide_* modes, each value is a list of prompt lists (one per word).
For ignore_* modes, each value is a flat list of prompts. Examples
Get prompts for ColorMNIST with number guidance
from laft.prompts import get_prompts
prompts = get_prompts("color_mnist", "guide_number")
# Access different prompt categories
print(f"Normal prompts: {len(prompts['normal'])} word groups")
print(f"Anomaly prompts: {len(prompts['anomaly'])} word groups")
# Example: First normal prompt group
print(prompts['normal'][0])
# Output: ['zero', 'a number zero', 'an image of zero', ...]
# Example: Using exact prompts (normal + anomaly)
for word_prompts in prompts['exact']:
print(f"{len(word_prompts)} variants for one word")
Get prompts for color guidance
prompts = get_prompts("color_mnist", "guide_color")
# Normal: red color variations
# Anomaly: green and blue color variations + auxiliary colors
print(prompts['normal'][0]) # ['red', 'a number red', ...]
print(prompts['anomaly'][0]) # ['green', 'a number green', ...]
Using ignore mode
# Ignore mode returns flattened prompt lists
prompts = get_prompts("color_mnist", "ignore_color")
# All prompts are flattened into single lists
print(type(prompts['normal'])) # list
print(prompts['normal'][:3])
# Output: ['zero', 'a number zero', 'an image of zero']
Working with different datasets
# Waterbirds dataset
waterbirds_prompts = get_prompts("waterbirds", "guide_background")
# CelebA dataset
celeba_prompts = get_prompts("celeba", "guide_hair_color")
get_labels
Extracts attend and ignore labels from attribute tensors based on guidance type.
from laft.prompts import get_labels
attend_name, ignore_name, attend_labels, ignore_labels = get_labels(
dataset_name="color_mnist",
attrs=dataset.attrs,
guidance="guide_number"
)
Parameters
Name of the dataset.Supported: color_mnist, waterbirds, celeba
Attribute tensor from the dataset with shape [num_samples, num_attrs].
Boolean values where False = normal, True = anomaly.
Type of guidance (e.g., "guide_number", "ignore_color").
Returns
Name of the attribute to attend to (e.g., "number" or "color").
Name of the attribute to ignore.
Boolean tensor of shape [num_samples] with labels for the attended attribute.
Boolean tensor of shape [num_samples] with labels for the ignored attribute.
Examples
from laft.datasets import build_semantic_dataset
from laft.prompts import get_labels
# Load dataset
dataset = build_semantic_dataset("color_mnist", "test")
# Get labels
attend_name, ignore_name, attend_labels, ignore_labels = get_labels(
dataset_name="color_mnist",
attrs=dataset.attrs,
guidance="guide_number"
)
print(f"Attending to: {attend_name}") # "number"
print(f"Ignoring: {ignore_name}") # "color"
print(f"Attend labels shape: {attend_labels.shape}") # [num_samples]
print(f"Number of anomalies in attended attribute: {attend_labels.sum()}")
Analyze label distribution
import torch
attend_name, ignore_name, attend_labels, ignore_labels = get_labels(
"color_mnist", dataset.attrs, "guide_color"
)
# Count samples by label combination
normal_normal = (~attend_labels & ~ignore_labels).sum()
normal_anomaly = (~attend_labels & ignore_labels).sum()
anomaly_normal = (attend_labels & ~ignore_labels).sum()
anomaly_anomaly = (attend_labels & ignore_labels).sum()
print(f"Normal {attend_name}, Normal {ignore_name}: {normal_normal}")
print(f"Normal {attend_name}, Anomaly {ignore_name}: {normal_anomaly}")
print(f"Anomaly {attend_name}, Normal {ignore_name}: {anomaly_normal}")
print(f"Anomaly {attend_name}, Anomaly {ignore_name}: {anomaly_anomaly}")
get_words
Returns word lists for a specific dataset and guidance type (without templates).
from laft.prompts import get_words
words = get_words(
dataset_name="color_mnist",
guidance="guide_number"
)
Parameters
Name of the dataset.Supported: color_mnist, waterbirds, celeba
Type of guidance (e.g., "guide_number", "ignore_color").
Returns
Dictionary mapping word categories to lists of words.Keys: normal, anomaly, half, exact, allFor guide_* modes: values are lists of words.
For ignore_* modes: values are flattened lists.
Examples
Get words for number guidance
from laft.prompts import get_words
words = get_words("color_mnist", "guide_number")
print(words['normal'])
# Output: ['zero', '0', 'one', '1', 'two', '2', 'three', '3', 'four', '4']
print(words['anomaly'])
# Output: ['five', '5', 'six', '6', 'seven', '7', 'eight', '8', 'nine', '9',
# 'ten', '10', 'eleven', '11', ...]
print(words['exact'])
# Output: ['zero', '0', ..., 'four', '4', 'five', '5', ..., 'nine', '9']
Get words for color guidance
words = get_words("color_mnist", "guide_color")
print(words['normal'])
# Output: ['red', 'ruby', 'scarlet', 'crimson', 'maroon', 'carmine', 'vermilion']
print(words['anomaly'])
# Output: ['green', 'lime', 'olive', 'jade', 'blue', 'azure', 'sky blue', 'navy',
# 'yellow', 'gold', 'amber', ...]
Compare words vs prompts
from laft.prompts import get_words, get_prompts
words = get_words("color_mnist", "guide_number")
prompts = get_prompts("color_mnist", "guide_number")
# Words are the base terms
print(f"Number of normal words: {len(words['normal'])}")
# Output: 10 (zero, 0, one, 1, ...)
# Prompts apply templates to each word
print(f"Number of normal prompt groups: {len(prompts['normal'])}")
# Output: 10 (one group per word)
print(f"Prompts per word: {len(prompts['normal'][0])}")
# Output: 19 (number of templates)
Prompt Structure
The prompt system uses templates to generate variations of each word:
TEMPLATES = [
"{}",
"a number {}",
"an image of {}",
"a picture of {}",
"a photo of {}",
"a drawing of {}",
"a sketch of {}",
"a figure of {}",
"{} letter",
"a number {} letter",
# ... and more
]
from laft.prompts import get_prompts
prompts = get_prompts("color_mnist", "guide_number")
# First word is "zero"
zero_prompts = prompts['normal'][0]
print(zero_prompts)
# ['zero', 'a number zero', 'an image of zero', 'a picture of zero', ...]
Guidance Modes
- guide_*: Returns structured prompts (list of lists) for each word
- ignore_*: Returns flattened prompts (single list) for token-level matching
# Guide mode: structured
guide_prompts = get_prompts("color_mnist", "guide_number")
print(type(guide_prompts['normal'])) # list of lists
# Ignore mode: flattened
ignore_prompts = get_prompts("color_mnist", "ignore_number")
print(type(ignore_prompts['normal'])) # flat list
