The Garbage In, Garbage Out Rule π§Ή
Training on thousands of low-quality responses causes models to output repetitive or unhelpful text. Recent studies prove that fine-tuning on 1,000 highly curated, clean instruction pairs produces superior alignments than 100,000 messy forum scrape records. We format instruction sets in standard ChatML or Llama-3 instruction templates to fit model tokenizers.
ChatML Formatting: Explicit roles (system, user, assistant) are defined using special tokens (e.g.
<|im_start|>) to prevent target models from confusing inputs during gradient calculation.