Self-training: why it works and how to avoid collapse

Self-Training on Synthetic Data: Mechanisms of Success and the Threshold of Model Collapse

Synthetic data is generated by models themselves for further training, addressing the shortage of real-world examples. Self-training works by formatting data into an optimal form, increasing the density of useful signals, and distilling skills from a teacher model to a student model. However, recursive use leads to model collapse—a narrowing of distribution and loss of rare events.

Synthetic data is created by algorithms: self-instruct for instructions, teacher-student distillation, generating tasks with solutions, and expanding datasets with similar examples. Unlike simulations, the focus is on generative models.

Advantages of synthetic data:

• Structured format: clear chain-of-thought, unified Q&A templates.
• High density: minimal noise, maximum diversity in phrasing a single concept.
• Pattern transfer: a strong model passes on solution logic, not just copying text.

The Risks of Recursion: How Model Collapse Occurs

In a closed loop of generation → training, the process repeats, mimicking the copying of degrading copies. Rare details are erased, errors accumulate, and the distribution narrows toward the mode.

Model collapse manifests gradually:

• Shrinking tails: rare cases disappear.
• Templatization: responses become homogeneous, losing nuance.
• False confidence: metrics on synthetic data improve, while those on real data decline.

Inflection Points: When Self-Training Spins Out of Control

Danger arises in three scenarios:

Scenario 1: Dominance of synthetic data. Real data must remain an anchor—at least 20–30% of the volume. If synthetic data exceeds 70%, the connection to reality breaks.

Scenario 2: Lack of verification. Without gold sets and hold-out data tests, the model optimizes for an illusion of quality.

Scenario 3: Non-verifiable tasks. Synthetic data is reliable for code (unit tests), mathematics (exact match), and logic (rules-based checks). In evaluative domains, there's a risk of self-confirmation.

| Task | Verifiability | Collapse Risk |

|--------|------------------|---------------|

| Code | High (tests) | Low |

| Mathematics | High | Low |

| Facts/judgments | Low | High |

Diagnosing Degradation in Production

Monitor symptoms on validation sets:

• Decrease in perplexity on the tails of long-tail distributions.
• Increase in templatization: cosine similarity of responses >0.8.
• Divergence of metrics: internal loss decreases, external benchmarks increase.

Use A/B tests: compare models with/without fresh synthetic data against a baseline on real tasks.

Practical Strategies for Safe Self-Training

Synthetic data is effective as a supplement. Key rules:

• Fixed real anchor: 30–50% real data in each batch, with rotation.
• Strict filtering:

- Perplexity filter < threshold.

- Deduplication (MinHash, Jaccard >0.9).

- Diversity sampling by embedding clusters.

• Targeted generation: focus on gaps—rare classes, edge cases, counterfactuals.
• Automatic verification:

- Code: lint + tests.

- Math: sympy/eval.

- Facts: internal RAG-check.

Check effectiveness via the triad:

• Utility: Δaccuracy on hold-out data.
• Robustness: tail loss < baseline.
• Cost: quality gain per GPU-hour.

Key Takeaways

• Synthetic data accelerates training but requires a real anchor for stability.
• Model collapse begins at >70% synthetic data without verification.
• Filtering and targeted generation reduce risks by 40–60% in benchmarks.
• Verifiable tasks (code, math) tolerate up to 90% synthetic data.

— Editorial Team