thealper2/gemma-3-270m-foodextract

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Sep 14, 2026Architecture:Transformer Featherless Exclusive Cold

Thealper2/gemma-3-270m-foodextract is a 0.3 billion parameter Gemma-3 model fine-tuned for structured food and drink information extraction from text descriptions. This model excels at parsing natural language input to identify and categorize food items, drinks, and associated tags. It is specifically optimized for tasks requiring precise extraction of culinary details, making it suitable for applications like recipe analysis or dietary tracking.

Loading preview...

Overview

This model, thealper2/gemma-3-270m-foodextract, is a specialized fine-tuned version of unsloth/gemma-3-270m-it. It has been specifically trained to extract structured food and drink information from natural language descriptions. The training utilized a LoRA (merged) SFT approach on the mrdbourke/FoodExtract-135k dataset, focusing on generating a consistent output format.

Key Capabilities

  • Structured Data Extraction: Converts free-form text descriptions into a structured JSON-like output, including food_or_drink status, tags, foods, and drinks.
  • Specific Tagging: Identifies and applies relevant tags such as nutrition panel, ingredient list, menu, recipe, food items, drink items, food advertisement, and food packaging.
  • Compact Size: Built on a 0.3 billion parameter Gemma-3 base, offering efficient performance for its specialized task.
  • High Accuracy on Specific Metrics: Achieves 0.9540 food_or_drink accuracy and 0.9118 Tags micro-F1 on its test set.

Good For

  • Food Information Processing: Ideal for applications requiring the automated extraction of food and drink details from various text sources.
  • Recipe Analysis: Can be used to parse recipes or food descriptions to identify ingredients and dishes.
  • Dietary Tracking: Useful for systems that need to categorize and list food and drink items from user input.

Limitations

  • The model reproduces labels generated by gpt-oss-120b, including potential errors from that source.
  • Item lists are comma-separated, which can lead to ambiguity if an item itself contains a comma.
  • Primarily trained on English descriptions and inputs longer than 1024 tokens were not seen during training.
  • Training was conducted on a deterministic subset of 20,000 examples from the cleaned training split.

Finetunes
1 models