Plain-English Summary
Creating instruction-following training data is expensive because it requires humans to write thousands of diverse prompts and ideal responses. Self-Instruct sidesteps this cost by having the language model generate its own training data. Starting with a small seed set of 175 human-written tasks, the model generates new instructions, classifies them, and produces corresponding responses. After filtering, this synthetic data is used to fine-tune the model itself.
The approach achieved 52% improvement on Super-NaturalInstructions compared to the base model, approaching the performance of InstructGPT which used extensive human annotation. This proved that models could largely teach themselves to follow instructions.
Key Innovation
The bootstrapping pipeline demonstrated that a capable language model contains sufficient knowledge to generate its own alignment data. The process iterates: generate diverse instructions, produce responses, filter for quality, and fine-tune. Each iteration improves the model’s instruction-following, which in turn improves the quality of generated data for the next iteration.
Quality filtering was critical — not all generated examples are useful. The paper developed filtering heuristics that retained diverse, high-quality examples while discarding redundant or low-quality ones.
Impact on the Field
Self-Instruct made instruction tuning accessible to anyone with API access to a capable model. Stanford’s Alpaca applied the technique to LLaMA using GPT-3.5 as the generator, producing a competitive instruction-following model for under $600. This triggered an explosion of cheaply produced instruction-tuned models that democratized access to ChatGPT-like capabilities.
The paper established synthetic data generation as a legitimate training strategy, paving the way for the “model distillation” approach that now dominates open-source model training.
Models That Built on This
Alpaca used GPT-3.5 to generate 52K instruction-following examples for LLaMA. Vicuna trained on user conversations shared from ChatGPT. WizardLM evolved instructions through increasing complexity. Orca used detailed chain-of-thought reasoning from GPT-4. The entire ecosystem of cheaply fine-tuned open-source chat models traces back to Self-Instruct’s demonstration that synthetic data could substitute for expensive human annotation.