Synthetic Data for Distillation

The most effective way to train a task-specific 8B parameter model is using synthetic data generated by a 70B+ parameter model.

Home / Guides / Synthetic Data Generation

The Distillation Pipeline

Instead of manually labeling 10,000 documents, write a highly detailed prompt for GPT-4 or Claude 3 Opus to annotate 100 documents. Verify the quality, then run the pipeline over the remaining 9,900 documents. Use this resulting dataset to fine-tune a local Llama-3-8B.

Avoiding Model Collapse

Ensure the seed data fed into the teacher model is highly diverse. Synthetic data generation works best for classification, extraction, and formatting tasks, rather than generating novel facts.

Internal Resources