Custom LLMs
Custom LLMs refer to Large Language Models that have been fine-tuned, trained from scratch, or adapted for specific domains, tasks, or user preferences. Unlike general-purpose foundation models, custom LLMs are optimized for particular use cases, often requiring less computational power if the base model is small or if parameter-efficient fine-tuning techniques are employed.
Key Characteristics
- Specialization: Tailored to specific datasets (e.g., medical, legal, code) or styles.
- Efficiency: Can run on consumer hardware if using small-language-models (SLMs) or quantization techniques.
- Privacy: Data remains local, avoiding cloud-based inference risks.
Training on Personal Hardware
Recent developments allow for the training and fine-tuning of small models on standard personal computers, democratizing access to model customization.
- Feasibility: It is possible to train custom LLMs on a PC without specialized high-end hardware, provided the model size is constrained (e.g., <7B parameters) and efficient training methods are used.
- Process: The workflow typically involves data preparation, selecting a base model, and using frameworks that support low-resource training.
- Reference: See Personal Computer Training of Small Language Models for Text Generation for a practical guide on training tiny LLMs in hours.