dksplit 0.3.1

High-performance string segmentation using BiLSTM-CRF

DKSplit

v0.3.1 — Model upgraded to EuroHPC infrastructure (Leonardo Booster, NVIDIA A100). ~3% accuracy improvement over v0.2.x on real-world domains. API unchanged.

String segmentation using BiLSTM-CRF. Splits concatenated words into meaningful parts.

DKSplit is a lightweight model trained on millions of labeled samples covering domain names, brand names, tech terms, and multilingual phrases. It uses a BiLSTM-CRF architecture (9.47M parameters) exported to ONNX with INT8 quantization, delivering fast CPU inference in a 9 MB package.

Originally built for domain name analysis at DomainKits, but works well on any concatenated text: hashtags, URLs, identifiers, compound strings.

Install

pip install dksplit

Usage

import dksplit

dksplit.split("chatgptlogin")
# ['chatgpt', 'login']

dksplit.split("kubernetescluster")
# ['kubernetes', 'cluster']

dksplit.split("mercibeaucoup")
# ['merci', 'beaucoup']

dksplit.split_batch(["openaikey", "microsoftoffice", "bitcoinprice"])
# [['openai', 'key'], ['microsoft', 'office'], ['bitcoin', 'price']]

What's New in v0.3.1

Model training upgraded from AWS to EuroHPC Leonardo Booster (NVIDIA A100), with optimized training configuration for better generalization. Improved accuracy on real-world domains, especially for brand names, multilingual inputs, and edge cases. The API is unchanged.

pip install --upgrade dksplit

Examples of improvements:

Input	v0.2.x	v0.3.1
cloudflarecdn	cloud flare cdn	cloudflare cdn
databricks	data bricks	databricks
instacart	insta cart	instacart
robinhood	robin hood	robinhood
mailchimp	mail chimp	mailchimp

Benchmark

Dataset

1,000 newly registered .com domains randomly sampled from ABTdomain.com daily feed (April 8, 2026). No filtering or cherry-picking. Ground truth was established through multi-model cross-validation (BiLSTM, Qwen 9B LoRA, Gemma 31B) and human audit.

The dataset and evaluation script are available on GitHub.

Results

Accuracy on 1,000 randomly sampled real-world .com domains, human-audited ground truth:

Model	Accuracy
DKSplit v0.3.1	85.0%
DKSplit v0.2.x	82.8%
WordSegment	54.0%
WordNinja	46.1%

DKSplit outperforms WordSegment by 31 percentage points and WordNinja by 39 percentage points.

Note: The accuracy above is measured against a single reference segmentation. Domain names are inherently ambiguous. For example, tiantian5 could be tiantian 5 (Chinese compound name) or tian tian 5 (two separate syllables); noranite could be nora nite or an intact brand; pikahug could be pika hug or an intact brand name. Our audit found ~5% of test samples have multiple valid segmentations. Accounting for these, effective accuracy is closer to 90%.

Comparison

Input	DKSplit v0.3.1	WordSegment	WordNinja
chatgptprompts	chatgpt prompts	chat gpt prompts	chat gp t prompts
tensorflowserving	tensorflow serving	tensor flow serving	tensor flow serving
spotifywrapped	spotify wrapped	spot if y wrapped	spot if y wrapped
ethereumwallet	ethereum wallet	e there um wallet	e there um wallet
cloudflarecdn	cloudflare cdn	cloud flare cdn	cloud flare cd n
kubernetescluster	kubernetes cluster	ku bernet es cluster	ku berne tes cluster
hackathonwinners	hackathon winners	hackathon winners	hack a th on winners
whatsappstatus	whatsapp status	what sapp status	what s app status
drwatsonai	dr watson ai	dr watson a i	dr watson a i
escribirenvozalta	escribir en voz alta	escribir env oz alta	es crib ire nv oz alta
tuvasou	tu vas ou	tuva sou	tuva so u
candidiasenuncamais	candidiase nunca mais	candid iase nunca mais	can didi as e nun cama is
robertdeniro	robert de niro	robert deniro	robert deniro
mercibeaucoup	merci beaucoup	merci beaucoup	mer ci beau coup

How It Works

DKSplit treats segmentation as a sequence labeling task.

The training data includes:

• LLM-labeled domain name segmentations
• Brand names
• Personal name combinations
• Multilingual phrases (English, French, German, Spanish, and more)
• Tech product names and terminology

At inference, the BiLSTM runs as an INT8-quantized ONNX model and CRF decoding is performed in NumPy — no GPU required.

Features

• Brand-aware: Recognizes thousands of brands, tech products, and proper nouns
• Multilingual: Handles English, French, German, Spanish, and romanized text
• Lightweight: 9 MB model, minimal dependencies (numpy + onnxruntime)
• Offline: No API keys, no internet required

Limitations

• Characters: Only a-z and 0-9. Input is automatically lowercased.
• Max length: 64 characters.
• Script: Latin script only. Non-Latin scripts (汉字, かな, 한글, العربية) are not supported.
• Ambiguity: Some inputs are genuinely ambiguous. DKSplit optimizes for the most common interpretation.
• Rare languages: Accuracy is highest on English and major European languages.

Requirements

• Python >= 3.8
• numpy
• onnxruntime

Links

• Website: domainkits.com, ABTdomain.com
• GitHub: github.com/ABTdomain/dksplit
• PyPI: pypi.org/project/dksplit
• Issues: GitHub Issues

License

This project is licensed under the Apache License 2.0.

Please attribute as: DKsplit by ABTdomain

Acknowledgements

 

The v0.3.1 model was trained on the Leonardo Booster supercomputer at CINECA, Italy, with computing resources provided by the EuroHPC Joint Undertaking through the Playground Access program (project AIFAC_P02_281). We thank EuroHPC JU for enabling SMEs to explore new possibilities with world-class HPC infrastructure.