adodabele/FoodExtract-gemma-3-270m-fine-tune-v1

FoodExtract-gemma-3-270m-fine-tune-v1 Overview

This model, developed by adodabele, is a specialized language model built upon the Gemma 3 270M architecture. It is fine-tuned for the specific task of extracting structured food and drink information from raw text inputs. The model processes text to determine if it relates to food or drink, assigns relevant tags, and extracts lists of food and drink items.

Key Capabilities

• Food/Drink Classification: Determines if input text describes food or drink (binary classification).
• Tagging: Assigns one or more predefined tags (e.g., 'nutrition_panel', 'ingredient list', 'menu', 'food_items', 'drink_items') to the text for fast filtering.
• Item Extraction: Identifies and lists edible food-related items and drink-related items present in the text.

Training and Data

The model was trained using Supervised Fine-Tuning (SFT) with Hugging Face's TRL library. Its training dataset, FoodExtract-1k, comprises 1400 samples of raw text paired with structured JSON outputs. These structured outputs were generated by gpt-oss-120b, ensuring high-quality extraction examples. The model is optimized to output a condensed text format for efficiency, with helper functions provided for converting between condensed and uncondensed structured data.

Ideal Use Cases

This model is particularly well-suited for:

• Filtering large text datasets: Efficiently sifting through extensive text collections, such as image captions (e.g., DataComp-1B), to identify and categorize food and drink related content.
• Structured data extraction: Automating the process of extracting specific food and drink entities and their attributes from unstructured text.