How We Train LLMs
At Liberty Apps, we develop large language models (LLMs) with a strong focus on accuracy, safety, and practical value. Our training pipeline combines cutting-edge machine learning techniques, powerful infrastructure, and rigorous evaluation standards to ensure our models deliver reliable and context-aware outputs.
1. Data Collection and Curation
Everything starts with data. We gather diverse, high-quality datasets that reflect how people communicate across domains—science, business, customer support, creative writing, and more. Key steps include:
•Sourcing from publicly available data, licensed datasets, and proprietary sources (when permitted).
•Filtering out low-quality, offensive, or irrelevant content.
•Balancing the dataset to reduce bias and overrepresentation.
We prioritise transparency and ethical data use, avoiding personally identifiable information and respecting content licenses.
2. Pretraining the Model
Pretraining teaches the model the basics of language—syntax, semantics, grammar, and general world knowledge.
•We use transformer-based architectures, optimised for large-scale learning.
•Training runs on distributed GPU/TPU clusters, allowing us to scale to billions of parameters.
•The model learns through self-supervised learning, predicting masked or next tokens in sequences.
This phase can take weeks, depending on model size and compute availability.
3. Fine-Tuning for Specific Use Cases
Once pretrained, the model undergoes fine-tuning—targeted training on curated datasets that reflect the desired use cases. This makes the model more helpful, safer, and aligned with user needs.
Examples of fine-tuning goals:
•Generating clean, structured responses
•Following instructions precisely
•Avoiding hallucinations and unsafe content
We also incorporate Reinforcement Learning from Human Feedback (RLHF) to refine behavior based on preferences collected from human reviewers.
4. Evaluation and Testing
We rigorously test our models using:
•Automated benchmarks (e.g., MMLU, HELM)
•Human evaluations for coherence, helpfulness, and safety
•Red teaming to expose vulnerabilities and edge cases
We continuously iterate based on evaluation results, updating training data, model parameters, and safety techniques.
5. Deployment and Monitoring
After passing internal thresholds, models are deployed via APIs, SDKs, or embedded into our products. We monitor usage in real-time to:
•Detect harmful or unintended outputs
•Collect anonymized feedback
•Improve performance over time
We remain committed to responsible AI deployment, including rate-limiting, abuse detection, and transparent documentation.
Our Commitment
Training LLMs isn’t just about scale—it’s about intent. At every stage, we prioritize:
• Safety
• Fairness
• Transparency
• Performance
We believe in building AI that earns trust and solves real problems—intelligently, reliably, and ethically.