A Dataset for Evaluating Large Language Models in Carbohydrate Estimation from Meal Descriptions

[Website] [Paper]

All data can be accessed in the data directory.

NutriBench consists of 5,000 human-verified meal descriptions with macro-nutrient labels, including carbohydrates, proteins, fats, and calories. It is divided into 15 subsets varying in meal description complexity.

The subsets are named according to the following format: {serving_unit}_{retrieval_type}_{number_of_food}_{number_of_serving}.csv.

Where:

• {serving_unit}: how the servings are measured in the description. {natural} (such as '1 cup') or {metric} (such as '50g').
• {retrieval_type}: to evaluate performance on RAG based methods we divided Nutribench in two. With the {direct} subsets, food items can be directly retrieved from a RAG DB with exact food name matches. With the {indirect} subsets there is no direct match between the queried food item name and the items in the RAG DB.
• {number_of_food}: the number of discrete food items in the description. One item {single}, two items {double}, or three items {triple}.
• {number_of_serving}: the number of servings in each description. One serving {single}, otherwise {multiple}.

For example:

natural_direct_single_multiple.csv indicates queries with natural serving unit, direct retrieval, a single food , and multiple servings of that single food.

# pip install pandas
import pandas as pd

file_path = 'data/metric_indirect_double_multiple.csv'
df = pd.read_csv(file_path)

For any questions, please contact the authors at {dongx1997,mdhaliwal}@ucsb.edu or [email protected]