Training configuration

Every fine-tuning job is defined by a JSON file matching the TrainingConfig interface.

Source: src/training/types.ts

Schema

type Framework = 'unsloth' | 'axolotl' | 'mlx';

interface TrainingConfig {
  framework:            Framework;
  baseModel:            string;   // Hugging Face repo or Ollama tag
  datasetPath:          string;   // .jsonl file (Alpaca or ShareGPT)
  outputDir:            string;   // where to save the adapter
  epochs:               number;
  batchSize:            number;
  gradientAccumulation: number;
  learningRate:         number;
  loraR:                number;
  loraAlpha:            number;
  maxSeqLength:         number;
}

The four required fields are framework, baseModel, datasetPath, outputDir. Everything else has a default.

Defaults

export const TRAINING_DEFAULTS = {
  epochs:               3,
  batchSize:            2,    // conservative for consumer VRAM
  gradientAccumulation: 4,
  learningRate:         2e-4,
  loraR:                8,    // smaller rank = less memory
  loraAlpha:            16,
  maxSeqLength:         2048,
};

These defaults target a 16 GB consumer GPU. Bump batchSize and loraR if you have more VRAM, lower them if you're hitting OOM.

Minimal example

{
  "framework": "unsloth",
  "baseModel": "meta-llama/Meta-Llama-3-8B",
  "datasetPath": "./data/alpaca.jsonl",
  "outputDir":   "./out/llama3-finetune"
}

datasetPath and outputDir may be relative — the CLI resolves them against the directory of the config file, not the current working directory. This keeps configs portable.

Full example

{
  "framework": "unsloth",
  "baseModel": "meta-llama/Meta-Llama-3-8B",
  "datasetPath": "./data/customer-support.jsonl",
  "outputDir":   "./out/support-bot",
  "epochs": 5,
  "batchSize": 4,
  "gradientAccumulation": 2,
  "learningRate": 1e-4,
  "loraR": 16,
  "loraAlpha": 32,
  "maxSeqLength": 4096
}

Submit it

locopilot train --config train.json          # Free — local in-process
locopilot train --config train.json --cloud  # Pro  — LocoPilot Cloud

See locopilot train for the full command surface.

Adapter selection

The framework you pick maps to a Python runner:

Framework	Runner	Best for
unsloth	src/training/adapters/unsloth_runner.py	Linux / Windows with NVIDIA GPU; ~2× faster than vanilla HF Trainer
axolotl	src/training/adapters/axolotl_runner.py	More configuration knobs; multi-GPU
mlx	src/training/adapters/mlx_runner.py	Apple Silicon (auto-selected on darwin/arm64)

See Adapters for installation and platform notes.

Output

When training finishes, outputDir contains:

out/llama3-finetune/
├── adapter_config.json
├── adapter_model.safetensors
├── tokenizer.json
└── training_args.json

The output_path field on the job row points at this directory.