A fine-tuning project for invoice extraction found that approximately 100 labeled invoices are sufficient to achieve 98% accuracy on known layouts using the Qwen2.5-1.5B-Instruct model with LoRA. Training on up to 1,000 invoices yielded only a marginal 1.6% improvement on these known layouts, while performance on unseen layouts plateaued around 77-81%, no better than using a prompt alone. Interestingly, training on a very small set of 25 invoices resulted in lower accuracy than using the prompt without any fine-tuning, suggesting that a minimal number of labeled examples should be included in the prompt rather than used for training. AI
IMPACT Fine-tuning efficiency for specific tasks is improved by understanding optimal data quantities, reducing wasted labeling effort.
RANK_REASON The item details a specific experiment and findings regarding fine-tuning a model for a particular task, including methodology and results. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
