unsloth-sft

Overview

Supervised Fine-Tuning (SFT) in Unsloth focuses on training models to follow instructions using specific formats. It provides tools for chat template mapping, multi-turn conversation synthesis via conversation_extension, and optimized dataset processing.

When to Use

• When training models on instruction-response datasets (e.g., Alpaca).
• When developing multi-turn conversational agents.
• When you need to standardize various dataset formats (ShareGPT, OpenAI) for training.

Decision Tree

1. Is your dataset single-turn?
   • Yes: Use conversation_extension to synthetically create multi-turn samples.
   • No: Map columns using standardize_sharegpt.
2. Are you training on Windows?
   • Yes: Set dataset_num_proc = 1 in SFTConfig.
   • No: Use multiple processes for faster mapping.
3. Want to increase multi-turn accuracy?
   • Yes: Enable masking of inputs to train on completions only.

Workflows