The SFT strategy is the standard, general-purpose fine-tuning approach using LoRA adapters.

Supervised Fine-Tuning uses labeled examples to teach the model new behaviors or knowledge through standard supervised learning.

✅ Simple and effective - Works for most use cases
✅ Well-tested - Industry standard approach
✅ Fast training - Faster than RLHF/DPO
✅ Low complexity - Easy to understand and debug
✅ Good quality - High-quality results for most tasks

{
  "strategy": "sft",
  "task": "text-generation",
  "model_name": "meta-llama/Llama-3.2-3B",
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.1
}

SFT uses LoRA (Low-Rank Adaptation) for efficient fine-tuning:

{
  "lora_r": 16,           // Rank (8, 16, 32, 64)
  "lora_alpha": 32,       // Alpha (usually 2x rank)
  "lora_dropout": 0.1,    // Dropout rate
  "target_modules": "all-linear"
}

Standard input-output pairs:

{"input": "Question or instruction", "output": "Expected response"}

See Dataset Formats for details.

• QLoRA Strategy - Memory-efficient variant
• Strategy Overview - Compare all strategies
• Configuration Guide - All options

SFT: The reliable standard for fine-tuning! ✨