Model Card — Leuks

1. Model Overview

Leuks is a family of large language models (LLMs) developed by Inferent. The models are designed to understand and generate natural language, assist with reasoning tasks, answer questions, write and debug code, summarize documents, and engage in multi-turn dialogue.

Leuks models are available through Inferent's first-party products (such as Inclick's AI Tutor) and through the Leuks API for developers and enterprise customers.

2. Intended Use

Leuks models are intended for:

• Education and tutoring assistance
• Content drafting, editing, and summarization
• Software development assistance (code generation, debugging, explanation)
• Research and information synthesis
• Business productivity and automation
• Creative writing and ideation
• Customer support and conversational interfaces

3. Out-of-Scope Uses

Leuks models are not designed for, and should not be used for, the following without appropriate human oversight, professional expertise, and legal authorization:

• Medical diagnosis or treatment: AI outputs are not a substitute for licensed medical professionals
• Legal advice: AI outputs do not constitute legal counsel
• Financial advice: AI outputs are not investment or financial recommendations
• Autonomous high-stakes decision making: AI outputs should not be used as the sole basis for consequential decisions affecting individuals
• Weapons design or dangerous synthesis: Absolutely prohibited
• CSAM or sexual content involving minors: Absolutely prohibited

4. Training Methodology

Leuks models are trained using a multi-stage process consistent with modern industry practices:

Pre-training Data

Models are pre-trained on a diverse mixture of text sources, including:

• Publicly available text from the internet, filtered for quality, safety, and relevance
• Digitized books and academic publications from licensed providers
• Code repositories available under open-source licenses
• Multilingual text to support multiple languages

Data is processed through quality and safety filters before use in training. We do not copy copyrighted works verbatim; our approach uses these materials to develop language understanding capabilities.

Supervised Fine-tuning

After pre-training, models are fine-tuned on curated examples of high-quality instruction-following behavior, generated by a combination of human annotators and internally generated synthetic data.

Reinforcement Learning from Human Feedback (RLHF)

Human raters evaluate model outputs and provide preference rankings. These rankings train a reward model, which is then used to further optimize the language model to produce outputs that humans find more helpful, accurate, and appropriate.

Safety Optimization

Dedicated safety training passes apply additional optimization to reduce harmful, biased, and misleading outputs. This includes adversarial training on refusal scenarios and harmful content classification.

5. Evaluation

Leuks models are evaluated across multiple dimensions before deployment:

• Capability benchmarks: Standard NLP and reasoning benchmarks to measure model performance
• Safety evaluations: Automated and human evaluation of model behavior on sensitive topics
• Bias audits: Evaluation of model outputs across demographic dimensions
• Red teaming: Adversarial testing to identify failure modes

Benchmark results are published in our Transparency Report when available.

6. Limitations

• Knowledge cutoff: Models have a training data cutoff date and may not reflect recent events or developments
• Reasoning: Models can make logical errors, especially on complex multi-step reasoning tasks
• Consistency: Outputs may vary across sessions for similar inputs
• Language: Performance may vary across languages; English is generally strongest
• Long contexts: Performance may degrade for very long inputs

7. Hallucination Disclaimer

Like all large language models, Leuks models may generate plausible-sounding but factually incorrect information—a phenomenon commonly referred to as "hallucination." Models may confidently state inaccurate facts, cite non-existent sources, or misremember details.

Users should never rely solely on Leuks model outputs for factual claims, especially in high-stakes contexts (medical, legal, financial, safety-critical decisions). Always verify important information from authoritative, primary sources.

8. Known Biases

Our models may reflect biases present in training data and human feedback, including:

• Western, English-language cultural perspectives may be overrepresented
• Historical biases in text data related to gender, race, religion, and other protected characteristics may be reflected in model outputs
• Annotator bias from the human feedback process

We actively work to identify and mitigate these biases. See our Responsible AI Policy for our mitigation approach.

9. Contact

For model-related inquiries or to report problematic outputs:
Email: [email protected]