#62 From NutriBench to Taralli: How far can you take a prompt?

I walk through how I used the NutriBench dataset and DSPy to rigorously evaluate and improve Taralli’s calorie estimation, moving beyond “vibes” with a clear Accuracy@10% metric. After testing multiple prompts, optimizers, and models, a few-shot setup with Gemini 3 Flash delivered about a 15% accuracy boost over the previous Gemini 2.5 Flash baseline, and I also shipped an offline, on-device LLM mode plus UI updates. I close with practical lessons: hill-climb with data, consider fine-tuning or RAG, and remember that good few-shot examples often win. Relevant links: Original article Taralli: home-cooked food tracking without the B.S. NutriBench paper (arXiv) NutriBench dataset viewer (Hugging Face) WWEIA overview (USDA) FAO/WHO GIFT individual food consumption NutriBench dataset (Hugging Face) NutriBench prompts (gist) Notebook with the process (gist) DSPy: BootstrapFewShotWithRandomSearch DSPy: MIPROv2 DSPy: GEPA overview Example prediction (gist) GEPA prompt example (gist) OpenRouter

There are no comments yet.

Please log in to write the first comment.