SemAxis

Interpretable NLI-based text features for scikit-learn.

SemAxis turns raw text into a feature matrix by asking an LLM to generate natural-language hypotheses and scoring each text against them with an NLI model. Every feature is a human-readable sentence — no black-box embeddings.

texts  ──►  LLM (hypothesis generation)  ──►  NLI scoring  ──►  X: (n_texts, n_features)

Both transformers are sklearn-compatible: they work inside Pipeline and are safe to use with cross_val_score.

Installation

pip install semaxis

Unsupervised

UnsupervisedTransformer generates hypotheses that characterize the text collection without labels.

from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
from semaxis import UnsupervisedTransformer

pipe = Pipeline([
    ("vect", UnsupervisedTransformer(
        llm="gpt-4o",
        nli_model="cross-encoder/nli-deberta-v3-large",
        n_features=20,
    )),
    ("clf", LogisticRegression()),
])

cross_val_score(pipe, texts, labels, cv=5)

Supervised

SupervisedTransformer generates hypotheses that discriminate between classes.

Binary

from semaxis import SupervisedTransformer

pipe = Pipeline([
    ("vect", SupervisedTransformer(
        llm="gpt-4o",
        nli_model="cross-encoder/nli-deberta-v3-large",
        n_features=20,
    )),
    ("clf", LogisticRegression()),
])

cross_val_score(pipe, texts, labels, cv=5)

Labels can be numeric (0/1) or strings ("positive"/"negative").

Multi-class

Use strategy="ovr" (one-vs-rest, default) or strategy="ovo" (one-vs-one):

# OvR: generates n_features hypotheses per class (k × n_features total)
vect = SupervisedTransformer(llm="gpt-4o", nli_model="...", n_features=10, strategy="ovr")

# OvO: generates n_features hypotheses per class pair (C(k,2) × n_features total)
vect = SupervisedTransformer(llm="gpt-4o", nli_model="...", n_features=10, strategy="ovo")

Interpreting features

After fitting, both transformers expose features_ (flat list of hypothesis strings). SupervisedTransformer also exposes feature_meta_, parallel to features_, which records which class pair each hypothesis came from.

vect = SupervisedTransformer(llm="gpt-4o", nli_model="...", n_features=5, strategy="ovr")
vect.fit(texts, labels)

for hypothesis, meta in zip(vect.features_, vect.feature_meta_):
    print(f"[{meta.positive} vs {meta.negative}]  {hypothesis}")

[cat vs rest]  This text describes feline behavior.
[cat vs rest]  This text mentions a cat or kitten.
[dog vs rest]  This text describes canine behavior.
[dog vs rest]  This text mentions a dog or puppy.
...

Combine with a linear model to get per-hypothesis coefficients:

from sklearn.linear_model import LogisticRegression
import numpy as np

X = vect.transform(texts)
clf = LogisticRegression().fit(X, labels)

for coef, hyp in sorted(zip(clf.coef_[0], vect.features_), key=lambda x: abs(x[0]), reverse=True):
    print(f"  {coef:+.3f}  {hyp}")

Custom LLM

For in-process inference, use LlamaCppClient backed by llama-cpp-python:

from llama_cpp import Llama
from semaxis import UnsupervisedTransformer
from semaxis import LlamaCppClient

llm = LlamaCppClient(Llama(model_path="path/to/model.gguf", n_ctx=4096))

vect = UnsupervisedTransformer(llm=llm, nli_model="cross-encoder/nli-deberta-v3-large")

Install the optional dependency with:

pip install "semaxis[llamacpp]"

Related Work

• Balek et al. (2025) — LLM-based feature generation for interpretable ML
• Yin et al. (2019) — NLI as zero-shot text classifier
• LogiPart (2025) — LLM hypothesis generation + NLI propagation

Citation

@software{semaxis2026,
  title  = {SemAxis: Interpretable NLI-based text features for scikit-learn},
  year   = {2026},
}