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

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

This model, developed by arjunraim, is a specialized language model based on the Gemma 3 270M architecture. It is fine-tuned for the precise task of extracting structured food and drink information from arbitrary text inputs. The model processes text to perform three key functions: classifying whether the text pertains to food or drink, assigning one or more predefined tags (e.g., 'ingredient list', 'menu', 'food items'), and extracting specific lists of food and drink items mentioned.

Key Capabilities

• Food/Drink Classification: Determines if input text describes food or drink (binary output).
• Tagging: Applies relevant tags from a predefined dictionary (nutrition_panel, ingredient list, menu, recipe, food_items, drink_items, food_advertisement, food_packaging) for quick filtering.
• Item Extraction: Identifies and lists all edible food-related and drink-related items present in the text.
• Efficient Output: Generates a condensed, non-JSON output format to minimize token generation, with helper functions provided for conversion to and from structured dictionaries.

Training and Use Cases

The model was trained using Supervised Fine-Tuning (SFT) with Hugging Face's TRL library on the FoodExtract-1k dataset, which contains 1400 samples of raw text and corresponding structured extractions. Its primary application is filtering extensive text datasets, such as image caption repositories like DataComp-1B, to isolate and categorize food and drink-related entries efficiently.