Part 7/13:
- Better alternatives exist: Prompt engineering, retrieval-augmented generation (RAG), and traditional algorithms with embeddings can often provide sufficient results with less overhead.
For instance, in complex procedural questions that demand aggregating information from multiple document sections, I the speaker notes that even large fine-tuned models sometimes struggle, producing accuracy no better than 62-75%. This demonstrates that simpler, more transparent approaches may be preferable initially.
Essential Steps Before Fine-Tuning
The speaker advocates a careful sequence:
- Start with prompt engineering: Design prompts and benchmarks to evaluate initial performance.