Self-Instruct, Evol-Instruct, and model collapse
When human demonstrations are scarce, models generate their own training data: seed prompts are expanded (Self-Instruct) or evolved into harder variants (Evol-Instruct), answered, filtered, and fed back. Without fresh external grounding, repeated self-training degrades diversity — model collapse.
Self-Instruct Loop
Seed prompts → LLM generates new prompts + responses → filter → add to dataset. Bootstrap coverage quickly from ~175 seed tasks.
Evol-Instruct
Each generation rewrites prompts to be more complex, constrained, or multi-step — WizardLM's recipe for escalating difficulty.
Collapse Guard
Mix synthetic data with human/curated data each round, and track output entropy. Pure self-training rounds slowly shrink the distribution's tails.
Keeping a fraction of real data in every synthetic round bounds capability loss: a 10% forgetting rate with 30% replay retains ~93% of the original distribution.
- Orca and Phi-style curricula show synthetic data quality depends on reasoning traces, not just answers — generate the rationale, then grade it.
- Every synthetic round should end at a rejection gate (next scene); unfiltered self-training is how collapse sneaks in.