artifex 0.10.0

Small Language Model Inference, Fine-Tuning, Evaluation and Observability.

Artifex

Small Language Model Inference, Fine-Tuning and Observability. No GPU, no labeled data needed.

---

Artifex is a Python library for:

1. Using pre-trained task-specific Small Language Models on CPU
2. Fine-tuning them on CPU without any training data — just based on your instructions for the task at hand.
   How is it possible?

   Artifex generates synthetic training data on-the-fly based on your instructions, and uses this data to fine-tune Small Language Models for your specific task. This approach allows you to create effective models without the need for large labeled datasets.
3. Tracking model performance locally with built-in evaluation and monitoring tools.

Why Artifex?

Modern AI workflows are often

• Expensive (API usage, GPUs)
• Dependent on third-parties
• Data-hungry (require large labeled datasets)

Artifex changes that:

• Run tiny models locally on CPU (100M params, 500MB)
• Keep all data private (no API required)
• Generate synthetic data automatically for fine-tuning
• Fine-tune models for specific tasks (moderation, NER, classification, etc.)

Available Models & Tasks

At this time, Artifex supports the following models:

Task	Available Languages	Description	Default Model	How to use
Guardrail	English, German, Spanish	Flags unsafe, harmful, or off-topic messages.	tanaos/tanaos-guardrail-v2 (English version, see the page for the other languages)	Examples
Intent Classification	English	Classifies user messages into predefined intent categories.	tanaos/tanaos-intent-classifier-v1	Examples
Reranker	English	Ranks a list of items or search results based on relevance to a query.	cross-encoder/mmarco-mMiniLMv2-L12-H384-v1	Examples
Sentiment Analysis	English	Determines the sentiment (positive, negative, neutral) of a given text.	tanaos/tanaos-sentiment-analysis-v1	Examples
Emotion Detection	English	Identifies the emotion expressed in a given text.	tanaos/tanaos-emotion-detection-v1	Examples
Named Entity Recognition	English	Detects and classifies named entities in text (e.g., persons, organizations, locations).	tanaos/tanaos-NER-v1	Examples
Text Anonymization	English	Removes personally identifiable information (PII) from text.	tanaos/tanaos-text-anonymizer-v1	Examples
Spam Detection	English, German, Spanish, Italian	Identifies whether a message is spam or not.	tanaos/tanaos-spam-detection-v1 (English version, see the page for the other languages)	Examples
Topic Classification	English	Classifies text into predefined topics.	tanaos/tanaos-topic-classification-v1	Examples
Text Summarization	English	Rewrites text in a more concise way.	tanaos/tanaos-text-summarization-v1	Examples

Looking for models in other languages? Our Enterprise License includes models in any language. Reach out at [e-mail] for more details.

For each model, Artifex provides:

1. Inference API to use a default, pre-trained Small Language Model to perform that task out-of-the-box locally on CPU.
2. Fine-tune API to fine-tune the default model based on your requirements, without any training data and on CPU.
3. Load API to load your fine-tuned model locally on CPU, and use it for inference or further fine-tuning.
4. Built-in, automatic evaluation and monitoring tools to track model performance over time, locally on your machine.

Quick Start

Install Artifex with:

pip install artifex

Guardrail Model

Use the default Guardrail model

Use Artifex's default guardrail model, which is trained to flag unsafe or harmful messages out-of-the-box:

from artifex import Artifex

guardrail = Artifex().guardrail()
print(guardrail("How do I make a bomb?"))

# >>> [{'is_safe': False, 'scores': {'violence': 0.625, 'non_violent_unethical': 0.0066, 'hate_speech': 0.0082, 'financial_crime': 0.0072, 'discrimination': 0.0029, 'drug_weapons': 0.6633, 'self_harm': 0.0109, 'privacy': 0.003, 'sexual_content': 0.0029, 'child_abuse': 0.005, 'terrorism_organized_crime': 0.1278, 'hacking': 0.0096, 'animal_abuse': 0.009, 'jailbreak_prompt_inj': 0.0131}}]

Learn more about the default guardrail model and what it considers safe vs unsafe on our Guardrail HF model page.

Create & use a custom Guardrail model

Need more control over what is considered safe vs unsafe? Fine-tune your own guardrail model, use it locally on CPU and keep it forever:

from artifex import Artifex

guardrail = Artifex().guardrail()

model_output_path = "./output_model/"

guardrail.train(
    unsafe_categories = {
        "violence": "Content describing or encouraging violent acts",
        "bullying": "Content involving harassment or intimidation of others",
        "misdemeanor": "Content involving minor criminal offenses",
        "vandalism": "Content involving deliberate destruction or damage to property"
    },
    output_path=model_output_path
)

guardrail.load(model_output_path)
print(guardrail("I want to destroy public property."))

# >>> [{'is_safe': False, 'scores': {'violence': 0.592, 'bullying': 0.0066, 'misdemeanor': 0.672, 'vandalism': 0.772}}]

Reranker model

Use the default Reranker model

Use Artifex's default reranker model, which is trained to rank items based on relevance out-of-the-box:

from artifex import Artifex

reranker = Artifex().reranker()

print(reranker(
    query="Best programming language for data science",
    documents=[
        "Java is a versatile language typically used for building large-scale applications.",
        "Python is widely used for data science due to its simplicity and extensive libraries.",
        "JavaScript is primarily used for web development.",
    ]
))

# >>> [('Python is widely used for data science due to its simplicity and extensive libraries.', 3.8346), ('Java is a versatile language typically used for building large-scale applications.', -0.8301), ('JavaScript is primarily used for web development.', -1.3784)]

Create & use a custom Reranker model

Want to fine-tune the Reranker model on a specific domain for better accuracy? Fine-tune your own reranker model, use it locally on CPU and keep it forever:

from artifex import Artifex

reranker = Artifex().reranker()

model_output_path = "./output_model/"

reranker.train(
    domain="e-commerce product search",
    output_path=model_output_path
)

reranker.load(model_output_path)
print(reranker(
    query="Laptop with long battery life",
    documents=[
        "A powerful gaming laptop with high-end graphics and performance.",
        "An affordable laptop suitable for basic tasks and web browsing.",
        "This laptop features a battery life of up to 12 hours, perfect for all-day use.",
    ]
))

# >>> [('This laptop features a battery life of up to 12 hours, perfect for all-day use.', 4.7381), ('A powerful gaming laptop with high-end graphics and performance.', -1.8824), ('An affordable laptop suitable for basic tasks and web browsing.', -2.7585)]

Other Tasks

For more details and examples on how to use Artifex for the other available tasks, check out our Documentation.

License, Paid & Enterprise solutions

Artifex is fair code distributed under the Sustainable Use License and Tanaos Enterprise License.

• Source available: source code is always visible
• Extensible: you can add your own models and functionalities

Enterprise licenses are available for additional features and support. Contact us at [e-mail] for more details. Enterprise features include:

• Higher-Performance Models
  • Improved accuracy
  • Better handling of edge cases
  • Reduced false positives/negatives
• Custom Models
  • Models fine-tuned on your specific data and requirements
  • Support for any language, domain or task
• Production-Ready Models
  • Models trained on 1000x more data
  • 10x lower inference latency
• Dedicated Support
  • Priority support
  • Custom feature requests
  • Dedicated onboarding and training

Additional information about license can be found in the docs.

Contributing

Contributions are welcome! Whether it's a bug fix or a new feature you want to add, we'd love your help. Check out our Contribution Guidelines to get started.

Documentation & Support

• Full documentation: https://docs.tanaos.com/artifex
• Get in touch: [e-mail]