Skip to main content

Models

High-level overview of the Models system: strategy, architecture, and quick commands.

Vision & strategy​

The Models system provides a config-driven, CLI-first framework for fine-tuning, managing, and deploying custom models.

  • Configuration-driven workflows (JSON/YAML)
  • Method-agnostic: LoRA, QLoRA, full fine-tune, adapters, prefix tuning
  • Dataset-centric: tools to create and validate datasets from your data
  • Production-ready: versioning, evaluation, deployment, monitoring

Architecture overview​

models/
├── methods/              # Fine-tuning methods (lora, qlora, full_finetune, adapters, ...)
├── datasets/             # Creation, formats, quality control, augmentation
├── config_examples/      # Example configurations
├── registry/             # Model & adapter registry
├── evaluation/           # Benchmarks, metrics, QA
├── deployment/           # Production deployment configs
└── utils/                # Helpers & tests

Core CLI​

# Dataset creation
llamafarm models create-dataset --source ./rag/data --format alpaca --output ./datasets/rag_qa.json

# Training
llamafarm models train --config ./config_examples/lora_basic.json --dataset ./datasets/rag_qa.json

# Evaluation
llamafarm models evaluate --model ./trained_models/lora_v1 --benchmark hellaswag

# Deployment
llamafarm models deploy --model ./trained_models/lora_v1 --target kubernetes --replicas 2

# Registry
llamafarm models list --type adapters
llamafarm models register --model ./trained_models/lora_v1 --name "rag_enhancement_v1"
llamafarm models rollback --name "rag_enhancement_v1" --version "1.0.0"

Method selection (quick guide)​

  • Limited GPU memory: prefer QLoRA or LoRA
  • Maximum quality: full fine-tuning with large datasets
  • Multi-domain: adapters/registry for hot-swapping

See also: Providers and Adapters pages for runtime integration details.