wtpsplit🪓

Segment any Text - Robustly, Efficiently, Adaptably⚡

This repository allows you to segment text into sentences or other semantic units. It implements the models from:

• SaT — Segment Any Text: A Universal Approach for Robust, Efficient and Adaptable Sentence Segmentation by Markus Frohmann, Igor Sterner, Benjamin Minixhofer, Ivan Vulić and Markus Schedl (state-of-the-art, encouraged).
• WtP — Where’s the Point? Self-Supervised Multilingual Punctuation-Agnostic Sentence Segmentation by Benjamin Minixhofer, Jonas Pfeiffer and Ivan Vulić (previous version, maintained for reproducibility).

The namesake WtP is maintained for consistency. Our new followup SaT provides robust, efficient and adaptable sentence segmentation across 85 languages at higher performance and less compute cost. Check out the state-of-the-art results in 8 distinct corpora and 85 languages demonstrated in our Segment any Text paper.

Installation

pip install wtpsplit

Or one of the following for ONNX support:

pip install wtpsplit[onnx-gpu]
pip install wtpsplit[onnx-cpu]

Usage

from wtpsplit import SaT

sat = SaT("sat-3l")
# optionally run on GPU for better performance
# also supports TPUs via e.g. sat.to("xla:0"), in that case pass `pad_last_batch=True` to sat.split
sat.half().to("cuda")

sat.split("This is a test This is another test.")
# returns ["This is a test ", "This is another test."]

# do this instead of calling sat.split on every text individually for much better performance
sat.split(["This is a test This is another test.", "And some more texts..."])
# returns an iterator yielding lists of sentences for every text

# use our '-sm' models for general sentence segmentation tasks
sat_sm = SaT("sat-3l-sm")
sat_sm.half().to("cuda") # optional, see above
sat_sm.split("this is a test this is another test")
# returns ["this is a test ", "this is another test"]

# use trained lora modules for strong adaptation to language & domain/style
sat_adapted = SaT("sat-3l", style_or_domain="ud", language="en")
sat_adapted.half().to("cuda") # optional, see above
sat_adapted.split("This is a test This is another test.")
# returns ['This is a test ', 'This is another test']

Faster PyTorch inference (torch.compile / TorchInductor)

For repeated inference without exporting to ONNX, compile the PyTorch model with TorchInductor via optimize() (PyTorch 2.0+). This is optional - the first split after optimize() can be slow while graphs are built. Call optimize() after to() / half() so the compiled graph matches your device and dtype.

from wtpsplit import SaT

sat = SaT("sat-3l-sm")
sat.half().to("cuda")
sat.optimize()  # backend="inductor" (default); dynamic shapes on

sat.split("This is a test This is another test.")

• SaT and WtP PyTorch checkpoints only — not available with ort_providers / ONNX.
• backend synonyms: "inductor", "torchinductor".
• Chunk length and the last batch size change between calls, so dynamic=True is the default. mode="reduce-overhead" (CUDA graphs) is faster only when every forward uses the same shapes; otherwise use the default or mode="max-autotune-no-cudagraphs".
• Optional on NVIDIA Ampere+: torch.set_float32_matmul_precision("high") before inference (faster fp32 matmuls).

NVIDIA AITune (auto backend selection, CUDA)

NVIDIA AITune benchmarks backends (TensorRT, Torch-TensorRT, Torch Inductor, …) on your GPU and picks a fast path. Requires Linux, CUDA, and a separate install:

pip install wtpsplit[aitune] --extra-index-url https://pypi.nvidia.com

sat = SaT("sat-3l-sm")
sat.half().to("cuda")
# first_wins tries TensorRT → Torch-TensorRT → Inductor when available
sat.optimize(backend="aitune")
# or only Inductor via AITune (faster tuning, no TensorRT dependency):
# sat.optimize(backend="aitune", aitune_strategy="inductor_only", aitune_max_batches=4)

sat.split("This is a test This is another test.")

ONNX Support

🚀 You can now enable even faster ONNX inference for sat and sat-sm models! 🚀

sat = SaT("sat-3l-sm", ort_providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

>>> from wtpsplit import SaT
>>> texts = ["This is a sentence. This is another sentence."] * 1000

# PyTorch GPU
>>> model_pytorch = SaT("sat-3l-sm")
>>> model_pytorch.half().to("cuda");
>>> %timeit list(model_pytorch.split(texts))
# 144 ms ± 252 μs per loop (mean ± std. dev. of 7 runs, 10 loops each)
# quite fast already, but...

# onnxruntime GPU
>>> model_ort = SaT("sat-3l-sm", ort_providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
>>> %timeit list(model_ort.split(texts))
# 94.9 ms ± 165 μs per loop (mean ± std. dev. of 7 runs, 10 loops each
# ...this should be ~50% faster! (tested on RTX 3090)

If you wish to use LoRA in combination with an ONNX model:

• Run scripts/export_to_onnx_sat.py with use_lora: True and an appropriate output_dir: <OUTPUT_DIR>.
  • If you have a local LoRA module, use lora_path.
  • If you wish to load a LoRA module from the HuggingFace hub, use style_or_domain and language.
• Load the ONNX model with merged LoRA weights: sat = SaT(<OUTPUT_DIR>, onnx_providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

Available Models

If you need a general sentence segmentation model, use -sm models (e.g., sat-3l-sm) For speed-sensitive applications, we recommend 3-layer models (sat-3l and sat-3l-sm). They provide a great tradeoff between speed and performance. The best models are our 12-layer models: sat-12l and sat-12l-sm.

The scores are macro-average F1 score across all available datasets for "English", and macro-average F1 score across all datasets and languages for "Multilingual". "adapted" means adapation via LoRA; check out the paper for details.

For comparison, here the English scores of some other tools:

Note that this library also supports previous WtP models. You can use them in essentially the same way as SaTmodels:

from wtpsplit import WtP

wtp = WtP("wtp-bert-mini")
# similar functionality as for SaT models
wtp.split("This is a test This is another test.")

For more details on WtP and reproduction details, see the WtP doc.

Paragraph Segmentation

Since SaT are trained to predict newline probablity, they can segment text into paragraphs in addition to sentences.

# returns a list of paragraphs, each containing a list of sentences
# adjust the paragraph threshold via the `paragraph_threshold` argument.
sat.split(text, do_paragraph_segmentation=True)

(NEW! v2.2+) Length-Constrained Segmentation

Control segment lengths with min_length and max_length parameters. This is useful when you need segments within specific size limits (e.g., for embedding models, storage, or downstream processing).

Basic Usage

from wtpsplit import SaT

sat = SaT("sat-3l-sm")

text = (
    "In the beginning God created the heaven and the earth. "
    "And the earth was without form, and void; and darkness was upon the face of the deep. "
    "And the Spirit of God moved upon the face of the waters. "
    "And God said, Let there be light: and there was light. "
    "And God saw the light, that it was good: and God divided the light from the darkness. "
    "And God called the light Day, and the darkness he called Night. "
    "And the evening and the morning were the first day."
)

# Split with a maximum segment length of 120 characters
segments = sat.split(text, max_length=120)
for i, s in enumerate(segments):
    print(f"[{len(s):3d} chars] {s}")
# [ 55 chars] In the beginning God created the heaven and the earth. 
# [ 86 chars] And the earth was without form, and void; and darkness was upon the face of the deep. 
# [112 chars] And the Spirit of God moved upon the face of the waters. And God said, Let there be light: and there was light. 
# [ 86 chars] And God saw the light, that it was good: and God divided the light from the darkness. 
# [115 chars] And God called the light Day, and the darkness he called Night. And the evening and the morning were the first day.

assert "".join(segments) == text  # text is perfectly preserved

# Enforce both min and max length
sat.split(text, min_length=80, max_length=200)

# Use the greedy algorithm for minimally faster (but less optimal) results
sat.split(text, max_length=120, algorithm="greedy")

Priors for Length Preference

Use priors to influence segment length distribution. Available priors:

# Gaussian prior (recommended): prefer segments around target_length
sat.split(text, max_length=100, prior_type="gaussian", 
          prior_kwargs={"target_length": 50, "spread": 10})

# Log-normal prior: right-skewed (more tolerant of longer segments)
sat.split(text, max_length=100, prior_type="lognormal", 
          prior_kwargs={"target_length": 70, "spread": 25})

# Clipped polynomial: hard cutoff at ±spread from target
sat.split(text, max_length=100, prior_type="clipped_polynomial", 
          prior_kwargs={"target_length": 60, "spread": 25})

Language-Aware Defaults

Pass lang_code to use language-specific defaults for target_length and spread (based on language-specific corpus statistics):

# German has longer average sentences → auto-uses target_length=90, spread=35
sat.split(text, max_length=150, prior_type="gaussian", 
          prior_kwargs={"lang_code": "de"})

# Chinese has shorter sentences → auto-uses target_length=45, spread=15
sat.split(text, max_length=100, prior_type="gaussian", 
          prior_kwargs={"lang_code": "zh"})

When using LoRA with a language, this happens automatically:

sat = SaT("sat-3l", style_or_domain="ud", language="de")
sat.split(text, max_length=150, prior_type="gaussian")  # auto-uses German defaults

How It Works

The Viterbi algorithm finds globally optimal segmentation points that balance:

• The model's sentence boundary predictions (where natural splits occur)
• Your length preferences (via the prior; if provided)

Text Reconstruction:

# With constraints (max_length or min_length):
original_text = "".join(segments)  # segments may contain newlines

# Without constraints (SaT default with split_on_input_newlines=True):
original_text = "\n".join(segments)

Note: When using length constraints, segments may contain newlines. If you want to remove them, you can just post-process the output.

Note: When max_length is set, the threshold parameter is ignored. The Viterbi/greedy algorithms use raw model probabilities directly instead of threshold-based filtering.

For more details, see the Length Constraints Documentation.

Adaptation

SaT can be domain- and style-adapted via LoRA. We provide trained LoRA modules for Universal Dependencies, OPUS100, Ersatz, and TED (i.e., ASR-style transcribed speeches) sentence styles in 81 languages for sat-3land sat-12l. Additionally, we provide LoRA modules for legal documents (laws and judgements) in 6 languages, code-switching in 4 language pairs, and tweets in 3 languages. For details, we refer to our paper.

We also provided verse segmentation modules for 16 genres for sat-12-no-limited-lookahead.

Load LoRA modules like this:

# requires both lang_code and style_or_domain
# for available ones, check the <model_repository>/loras folder
sat_lora = SaT("sat-3l", style_or_domain="ud", language="en")
sat_lora.split("Hello this is a test But this is different now Now the next one starts looool")
# now for a highly distinct domain
sat_lora_distinct = SaT("sat-12l", style_or_domain="code-switching", language="es-en")
sat_lora_distinct.split("in the morning over there cada vez que yo decía algo él me decía algo")

You can also freely adapt the segmentation threshold, with a higher threshold leading to more conservative segmentation:

sat.split("This is a test This is another test.", threshold=0.4)
# works similarly for lora; but thresholds are higher
sat_lora.split("Hello this is a test But this is different now Now the next one starts looool", threshold=0.7)

Advanced Usage

Get the newline or sentence boundary probabilities for a text:

# returns newline probabilities (supports batching!)
sat.predict_proba(text)

Load a SaT model in HuggingFace transformers:

# import library to register the custom models 
import wtpsplit.models
from transformers import AutoModelForTokenClassification

model = AutoModelForTokenClassification.from_pretrained("segment-any-text/sat-3l-sm") # or some other model name; see https://huggingface.co/segment-any-text

Adapt to your own corpus via LoRA

Our models can be efficiently adapted via LoRA in a powerful way. Only 10-100 training segmented training sentences should already improve performance considerably. To do so:

Clone the repository and install requirements:

git clone https://github.com/segment-any-text/wtpsplit
cd wtpsplit
pip install -r requirements.txt
pip install adapters==0.2.1 --no-dependencies
cd ..

1. Create data in this format:

import torch

torch.save(
    {
        "language_code": {
            "sentence": {
                "dummy-dataset": {
                    "meta": {
                        "train_data": ["train sentence 1", "train sentence 2"],
                    },
                    "data": [
                        "test sentence 1",
                        "test sentence 2",
                    ]
                }
            }
        }
    },
    "dummy-dataset.pth"
)

Note that there should not be any newlines within individual sentences! This now raises an error. Instead, each entry of a list should be a sentence, and there should be no "\n" characters. So your corpus should already be well-split.

2. Create/adapt config; provide base model via model_name_or_path and training data .pth via text_path:

configs/lora/lora_dummy_config.json

We recommend starting using this config, and adapting model_name_or_path, output_dir, and text_path if needed. You may also wish to adapt other aspects such as adapter_config and batch sizes, but this is more experimental.

3. Train LoRA:

python3 wtpsplit/train/train_lora.py configs/lora/lora_dummy_config.json

4. Once training is done, provide your saved module's path to SaT:

sat_lora_adapted = SaT("model-used", lora_path="dummy_lora_path")
sat_lora_adapted.split("Some domains-specific or styled text")

Important: Use the same model variant for inference as for training (e.g. sat-12l-sm and sat-12l have different configs; an adapter trained on one cannot be loaded on the other).

Adjust the dataset name, language and model in the above to your needs.

Reproducing the paper

configs/ contains the configs for the runs from the paper for base and sm models as well as LoRA modules. Launch training for each of them like this:

python3 wtpsplit/train/train.py configs/<config_name>.json
python3 wtpsplit/train/train_sm.py configs/<config_name>.json
python3 wtpsplit/train/train_lora.py configs/<config_name>.json

In addition:

• wtpsplit/data_acquisition contains the code for obtaining evaluation data and raw text from the mC4 corpus.
• wtpsplit/evaluation contains the code for:
  • evaluation (i.e. sentence segmentation results) via intrinsic.py.
  • short-sequence evaluation (i.e. sentence segmentation results for pairs/k-mers of sentences) via intrinsic_pairwise.py.
  • LLM baseline evaluation (llm_sentence.py), legal baseline evaluation (legal_baselines.py)
  • baseline (PySBD, nltk, etc.) evaluation results in intrinsic_baselines.py and intrinsic_baselines_multi.py
  • Raw results in JSON format are also in evaluation_results/
  • Statistical significane testing code and results ara in stat_tests/
  • punctuation annotation experiments in punct_annotation.py and punct_annotation_wtp.py (WtP only)
  • extrinsic evaluation on Machine Translation in extrinsic.py (WtP only)

Ensure to install packages from requirements.txt beforehand.

Supported Languages

Table with supported languages

iso	Name
af	Afrikaans
am	Amharic
ar	Arabic
az	Azerbaijani
be	Belarusian
bg	Bulgarian
bn	Bengali
ca	Catalan
ceb	Cebuano
cs	Czech
cy	Welsh
da	Danish
de	German
el	Greek
en	English
eo	Esperanto
es	Spanish
et	Estonian
eu	Basque
fa	Persian
fi	Finnish
fr	French
fy	Western Frisian
ga	Irish
gd	Scottish Gaelic
gl	Galician
gu	Gujarati
ha	Hausa
he	Hebrew
hi	Hindi
hu	Hungarian
hy	Armenian
id	Indonesian
ig	Igbo
is	Icelandic
it	Italian
ja	Japanese
jv	Javanese
ka	Georgian
kk	Kazakh
km	Central Khmer
kn	Kannada
ko	Korean
ku	Kurdish
ky	Kirghiz
la	Latin
lt	Lithuanian
lv	Latvian
mg	Malagasy
mk	Macedonian
ml	Malayalam
mn	Mongolian
mr	Marathi
ms	Malay
mt	Maltese
my	Burmese
ne	Nepali
nl	Dutch
no	Norwegian
pa	Panjabi
pl	Polish
ps	Pushto
pt	Portuguese
ro	Romanian
ru	Russian
si	Sinhala
sk	Slovak
sl	Slovenian
sq	Albanian
sr	Serbian
sv	Swedish
ta	Tamil
te	Telugu
tg	Tajik
th	Thai
tr	Turkish
uk	Ukrainian
ur	Urdu
uz	Uzbek
vi	Vietnamese
xh	Xhosa
yi	Yiddish
yo	Yoruba
zh	Chinese
zu	Zulu

For details, please see our Segment any Text paper.

Community Ports

• Rust: wtsplit-rs by @19h

Note: Community ports are independently maintained and may have different feature sets or update schedules.

Citations

For the SaT models, please kindly cite our paper:

@inproceedings{frohmann-etal-2024-segment,
    title = "Segment Any Text: A Universal Approach for Robust, Efficient and Adaptable Sentence Segmentation",
    author = "Frohmann, Markus  and
      Sterner, Igor  and
      Vuli{\'c}, Ivan  and
      Minixhofer, Benjamin  and
      Schedl, Markus",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.665",
    pages = "11908--11941"
}

For the library and the WtP models, please cite:

@inproceedings{minixhofer-etal-2023-wheres,
    title = "Where{'}s the Point? Self-Supervised Multilingual Punctuation-Agnostic Sentence Segmentation",
    author = "Minixhofer, Benjamin  and
      Pfeiffer, Jonas  and
      Vuli{\'c}, Ivan",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.398",
    pages = "7215--7235"
}

Acknowledgments

This research was funded in whole or in part by the Austrian Science Fund (FWF): P36413, P33526, and DFH-23, and by the State of Upper Austria and the Federal Ministry of Education, Science, and Research, through grants LIT-2021-YOU-215. In addition, Ivan Vulic and Benjamin Minixhofer have been supported through the Royal Society University Research Fellowship ‘Inclusive and Sustainable Language Technology for a Truly Multilingual World’ (no 221137) awarded to Ivan Vulić. This research has also been supported with Cloud TPUs from Google’s TPU Research Cloud (TRC). This work was also supported by compute credits from a Cohere For AI Research Grant, these grants are designed to support academic partners conducting research with the goal of releasing scientific artifacts and data for good projects. We also thank Simone Teufel for fruitful discussions.

---

For any questions, please create an issue or send an email to [email protected], and I will get back to you as soon as possible.