Scikit-learn friendly library to explain, predict, and steer text models/data.
Also a bunch of utilities for getting started with text data.

📖 demo notebooks

Explainable modeling/steering

📖Demo notebooks   🗂️ Doc   🔗 Reference code   📄 Research paper
⌛ We plan to support other interpretable algorithms like RLPrompt, CBMs, and NBDT. If you want to contribute an algorithm, feel free to open a PR 😄

General utilities

Quickstart

Installation: pip install imodelsx (or, for more control, clone and install from source)

import imodelsx.llm
llm = imodelsx.llm.get_llm('claude-code:haiku')  # or 'gpt-4o-mini', 'Qwen/Qwen2.5-0.5B-Instruct', ...
llm('What is the capital of Spain?')  # outputs are cached in ~/.CACHE_LLM
llm(['prompt 1', 'prompt 2'])  # API models run lists of prompts in parallel threads
llm = imodelsx.llm.get_llm('gpt-4o-mini', api_key='sk-...')  # pass a key directly

Demos: see the demo notebooks

Natural-language explanations

Tree-prompt

from imodelsx import TreePromptClassifier
import datasets
import numpy as np
from sklearn.tree import plot_tree
import matplotlib.pyplot as plt

# set up data
rng = np.random.default_rng(seed=42)
dset_train = datasets.load_dataset('rotten_tomatoes')['train']
dset_train = dset_train.select(rng.choice(
    len(dset_train), size=100, replace=False))
dset_val = datasets.load_dataset('rotten_tomatoes')['validation']
dset_val = dset_val.select(rng.choice(
    len(dset_val), size=100, replace=False))

# set up arguments
prompts = [
    "This movie is",
    " Positive or Negative? The movie was",
    " The sentiment of the movie was",
    " The plot of the movie was really",
    " The acting in the movie was",
]
verbalizer = {0: " Negative.", 1: " Positive."}
checkpoint = "gpt2"

# fit model
m = TreePromptClassifier(
    checkpoint=checkpoint,
    prompts=prompts,
    verbalizer=verbalizer,
    cache_prompt_features_dir=None,  # 'cache_prompt_features_dir/gp2',
)
m.fit(dset_train["text"], dset_train["label"])


# compute accuracy
preds = m.predict(dset_val['text'])
print('\nTree-Prompt acc (val) ->',
      np.mean(preds == dset_val['label']))  # -> 0.7

# compare to accuracy for individual prompts
for i, prompt in enumerate(prompts):
    print(i, prompt, '->', m.prompt_accs_[i])  # -> 0.65, 0.5, 0.5, 0.56, 0.51

# visualize decision tree
plot_tree(
    m.clf_,
    fontsize=10,
    feature_names=m.feature_names_,
    class_names=list(verbalizer.values()),
    filled=True,
)
plt.show()

iPrompt

from imodelsx import explain_dataset_iprompt, get_add_two_numbers_dataset

# get a simple dataset of adding two numbers
input_strings, output_strings = get_add_two_numbers_dataset(num_examples=100)
for i in range(5):
    print(repr(input_strings[i]), repr(output_strings[i]))

# explain the relationship between the inputs and outputs
# with a natural-language prompt string
prompts, metadata = explain_dataset_iprompt(
    input_strings=input_strings,
    output_strings=output_strings,
    checkpoint='EleutherAI/gpt-j-6B', # which language model to use
    num_learned_tokens=3, # how long of a prompt to learn
    n_shots=3, # shots per example
    n_epochs=15, # how many epochs to search
    verbose=0, # how much to print
    llm_float16=True, # whether to load the model in float_16
)
--------
prompts is a list of found natural-language prompt strings

D3 (DescribeDistributionalDifferences)

from imodelsx import explain_dataset_d3
hypotheses, hypothesis_scores = explain_dataset_d3(
    pos=positive_samples, # List[str] of positive examples
    neg=negative_samples, # another List[str]
    num_steps=100,
    num_folds=2,
    batch_size=64,
)

SASC

Here, we explain a module rather than a dataset

from imodelsx import explain_module_sasc
# a toy module that responds to the length of a string
mod = lambda str_list: np.array([len(s) for s in str_list])

# a toy dataset where the longest strings are animals
text_str_list = ["red", "blue", "x", "1", "2", "hippopotamus", "elephant", "rhinoceros"]
explanation_dict = explain_module_sasc(
    text_str_list,
    mod,
    ngrams=1,
)

Aug-imodels

Use these just a like a scikit-learn model. During training, they fit better features via LLMs, but at test-time they are extremely fast and completely transparent.

from imodelsx import AugLinearClassifier, AugTreeClassifier, AugLinearRegressor, AugTreeRegressor
import datasets
import numpy as np

# set up data
dset = datasets.load_dataset('rotten_tomatoes')['train']
dset = dset.select(np.random.choice(len(dset), size=300, replace=False))
dset_val = datasets.load_dataset('rotten_tomatoes')['validation']
dset_val = dset_val.select(np.random.choice(len(dset_val), size=300, replace=False))

# fit model
m = AugLinearClassifier(
    checkpoint='textattack/distilbert-base-uncased-rotten-tomatoes',
    ngrams=2, # use bigrams
)
m.fit(dset['text'], dset['label'])

# predict
preds = m.predict(dset_val['text'])
print('acc_val', np.mean(preds == dset_val['label']))

# interpret
print('Total ngram coefficients: ', len(m.coefs_dict_))
print('Most positive ngrams')
for k, v in sorted(m.coefs_dict_.items(), key=lambda item: item[1], reverse=True)[:8]:
    print('\t', k, round(v, 2))
print('Most negative ngrams')
for k, v in sorted(m.coefs_dict_.items(), key=lambda item: item[1])[:8]:
    print('\t', k, round(v, 2))

KAN

import imodelsx
from sklearn.datasets import make_classification, make_regression
from sklearn.metrics import accuracy_score
import numpy as np

X, y = make_classification(n_samples=5000, n_features=5, n_informative=3)
model = imodelsx.KANClassifier(hidden_layer_size=64, device='cpu',
                               regularize_activation=1.0, regularize_entropy=1.0)
model.fit(X, y)
y_pred = model.predict(X)
print('Test acc', accuracy_score(y, y_pred))

# now try regression
X, y = make_regression(n_samples=5000, n_features=5, n_informative=3)
model = imodelsx.kan.KANRegressor(hidden_layer_size=64, device='cpu',
                                  regularize_activation=1.0, regularize_entropy=1.0)
model.fit(X, y)
y_pred = model.predict(X)
print('Test correlation', np.corrcoef(y, y_pred.flatten())[0, 1])

General utilities

Easy baselines

Easy-to-fit baselines that follow the sklearn API.

from imodelsx import LinearFinetuneClassifier, LinearNgramClassifier
# fit a simple one-layer finetune on top of LLM embeddings
m = LinearFinetuneClassifier(
    checkpoint='distilbert-base-uncased',
)
m.fit(dset['text'], dset['label'])
preds = m.predict(dset_val['text'])
acc = (preds == dset_val['label']).mean()
print('validation acc', acc)

LLM wrapper

Easy API for calling different language models with caching (much more lightweight than langchain).

import imodelsx.llm
# supports huggingface, openai, anthropic, claude code, and codex checkpoints (see LLM providers in the quickstart)
llm = imodelsx.llm.get_llm(
    checkpoint="gpt2-xl",  # gpt-4o-mini, claude-haiku-5-5, claude-code:haiku, codex:gpt-5-codex, ...
    CACHE_DIR=".cache",
)
out = llm("May the Force be")
llm("May the Force be") # when computing the same string again, uses the cache

Data wrapper

API for loading huggingface datasets with basic preprocessing.

import imodelsx.data
dset, dataset_key_text = imodelsx.data.load_huggingface_dataset('ag_news')
# Ensures that dset has a split named 'train' and 'validation',
# and that the input data is contained for each split in a column given by {dataset_key_text}

Related work

• imodels package (JOSS 2021 github) - interpretable ML package for concise, transparent, and accurate predictive modeling (sklearn-compatible).
• Rethinking Interpretability in the Era of Large Language Models (arXiv 2024 pdf) - overview of using LLMs to interpret datasets and yield natural-language explanations
• Experiments in using clinical rule development: github
• Experiments in automatically generating brain explanations: github
• Interpretation regularization (ICML 2020 pdf, github) - penalizes CD / ACD scores during training to make models generalize better